add faq to notifications docs (#23939 )

Miscellaneous fixes (0.18 beta) (#23934 )
* add host npu requirements to docs * allow toggling live audio transcription via mqtt * improve spacing consistency on mobile drawers * fix clearing the region grid not surviving a restart
2026-08-10 21:01:10 +03:00 · 2026-08-08 11:12:13 -06:00 · 2026-08-08 07:20:09 -06:00 · 2026-08-08 06:09:12 -05:00 · 2026-08-08 06:09:12 -05:00 · 2026-08-08 06:09:12 -05:00
274 changed files with 12001 additions and 2038 deletions
@@ -8,6 +8,7 @@ amdgpu
 analyzeduration
 Annke
 apexcharts
+Aqara
 arange
 argmax
 argmin
@@ -64,6 +65,7 @@ dsize
 dtype
 ECONNRESET
 edgetpu
+Eufy
 facenet
 fastapi
 faststart
@@ -82,6 +84,7 @@ frontdoor
 fstype
 fullchain
 fullscreen
+gatekeep
 genai
 generativeai
 genpts
@@ -10,8 +10,11 @@ body:

        Before submitting, read the [beta documentation][docs].

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [docs]: https://docs-dev.frigate.video/
        [discussions]: https://github.com/blakeblackshear/frigate/discussions
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -8,9 +8,12 @@ body:

        Before submitting your support request, please [search the discussions][discussions], read the [official Frigate documentation][docs], and read the [Frigate FAQ][faq] pinned at the Discussion page to see if your question has already been answered by the community.

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -8,9 +8,12 @@ body:

        Before submitting your support request, please [search the discussions][discussions], read the [official Frigate documentation][docs], and read the [Frigate FAQ][faq] pinned at the Discussion page to see if your question has already been answered by the community.

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -8,9 +8,12 @@ body:

        Before submitting your support request, please [search the discussions][discussions], read the [official Frigate documentation][docs], and read the [Frigate FAQ][faq] pinned at the Discussion page to see if your question has already been answered by the community.

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -8,9 +8,12 @@ body:

        Before submitting your support request, please [search the discussions][discussions], read the [official Frigate documentation][docs], and read the [Frigate FAQ][faq] pinned at the Discussion page to see if your question has already been answered by the community.

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -8,9 +8,12 @@ body:

        Before submitting your support request, please [search the discussions][discussions], read the [official Frigate documentation][docs], and read the [Frigate FAQ][faq] pinned at the Discussion page to see if your question has already been answered by the community.

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -10,9 +10,12 @@ body:

        **If you are looking for support, start a new discussion and use a support category.**

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: textarea
    id: description
    attributes:
@@ -12,11 +12,14 @@ body:

        **If you are unsure if your issue is actually a bug or not, please submit a support request first.**

+        By posting here you agree to follow our [AI policy][ai-policy]. Posts that appear to be written by an AI on your behalf may be closed without a response.
+
        [discussions]: https://www.github.com/blakeblackshear/frigate/discussions
        [prs]: https://www.github.com/blakeblackshear/frigate/pulls
        [docs]: https://docs.frigate.video
        [faq]: https://github.com/blakeblackshear/frigate/discussions/12724
        [ai]: https://docs.frigate.video
+        [ai-policy]: https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
  - type: checkboxes
    attributes:
      label: Checklist
@@ -7,6 +7,13 @@ assignees: ''

 ---

+<!--
+  By posting here you agree to follow our AI policy:
+  https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md
+
+  Requests that appear to be written by an AI on your behalf may be closed without a response.
+-->
+
 **Describe what you are trying to accomplish and why in non technical terms**
 I want to be able to ... so that I can ... 

@@ -1,4 +1,4 @@
-_Please read the [contributing guidelines](https://github.com/blakeblackshear/frigate/blob/dev/CONTRIBUTING.md) before submitting a PR._
+_Please read the [contributing guidelines](https://github.com/blakeblackshear/frigate/blob/dev/CONTRIBUTING.md) and the [AI policy](https://github.com/blakeblackshear/frigate/blob/dev/AI_POLICY.md) before submitting a PR. Every PR must be read and submitted by a person, and PRs that appear to be unreviewed AI output will be closed without review._

 ## Proposed change

@@ -0,0 +1,126 @@
+# Frigate AI Policy
+
+## TL;DR
+
+- **Use AI tools if they help you.** We do too. This is about what you post, not which tools you use to write it.
+- **A person has to read it and send it.** Don't wire a bot or an agent up to post on your behalf.
+- **Write your posts yourself.** Your own words, the template filled in, and you answering maintainers rather than your assistant.
+- **Don't paste an AI's guess at the cause as though it were a diagnosis.** Tell us what you actually observed.
+- **Read your code before you submit it.** Disclose that AI was used, and be ready to explain every line.
+- **If we misjudge something you wrote, just say so.** We'll take you at your word.
+
+The rest of this document explains each of these, and why.
+
+## Scope
+
+AI tools are a reality of modern development and we're not opposed to their use. You are responsible for anything you submit, however it was produced, and we are responsible for anything we merge and release. We hold a high bar for both.
+
+This policy applies everywhere this project is discussed: issues, discussions, pull requests, code reviews, and commit comments.
+
+## Why this exists
+
+Frigate is built and supported by a small group of maintainers and a community of volunteers who read every post and review every pull request. Nobody here is paid to do it, and time spent reading a post is time not spent fixing bugs or building features.
+
+We're not opposed to AI tools. We use them too. But content generated by an AI and submitted without review costs a real person real time, and usually gives them less to work with than a few honest sentences would have. That is the problem this policy addresses.
+
+## A person has to be in the loop
+
+Every issue, discussion, comment, and pull request here must be read and submitted by a person. Using an AI tool to help you write is fine. Wiring one up to post on your behalf is not.
+
+Specifically, do not:
+
+- Connect a bot or agent to GitHub that opens issues, discussions, or pull requests without you reading them first
+- Post output from a tool you have not read
+- Use tooling to file bulk or drive-by contributions across the repository
+
+We will close anything we believe was posted without a person reading it, and we may mark it as spam. Posts that skip the templates are the most common sign of this.
+
+## Issues, discussions, and comments
+
+We do not mind if you use AI tools to help you write. Do not have tools post unreviewed content on your behalf. We may hide any comment we believe to be unreviewed AI output.
+
+Keep posts to what is needed to communicate your point. A long, confidently written, AI-padded post is harder to help with than a short direct one, not easier, and it is usually obvious.
+
+**Describe your actual problem in your own words.** Tell us what you did, what you expected, and what actually happened. That is the information we need, and only you have it.
+
+**Do not paste an AI's guess at the cause as though it were a diagnosis.** It is frequently wrong in ways that send everyone down the wrong path, and it buries the details that would have led to the real answer. We would rather see what you observed than what a model inferred.
+
+**Fill in the template completely.** The templates ask for logs, config, version, and hardware because those are the things needed to help you. An AI cannot supply them for you, and a post missing them cannot be acted on.
+
+**Answer maintainers yourself.** If we ask you a question, we are asking _you_, not your AI assistant. These are the spaces where we build trust and understanding with the community, and that only works if we're talking to each other. Using AI to fix your grammar or clarity is fine, but the substance has to be yours.
+
+This applies to pull request descriptions and review replies as much as it does to bug reports and discussions.
+
+### Quoting AI output
+
+If you want to include something an AI told you, it must be:
+
+- In a quote block, using `>`
+- Disclosed as AI output, saying which tool it came from
+- Accompanied by your own comment explaining why you think it is relevant
+
+Keep the excerpt short. Do not paste long transcripts.
+
+### Non-native English speakers
+
+AI is genuinely useful for participating in a project that operates in English, and we would rather hear from you through a translation tool than not hear from you at all. Using AI to improve the grammar or clarity of something you wrote yourself is fine.
+
+If you are translating your posts, make sure the translation says what you meant. Including your original text in a `<details>` block helps us verify the translation if something reads oddly, and keeps the thread readable.
+
+## Code contributions
+
+We need to understand your relationship with the code you're submitting. The more AI was involved, the more important it is that you've genuinely reviewed, tested, and understood what it produced.
+
+Because of the long-term maintenance burden every merged change creates, we require a human in the loop who understands the work the AI produced. Pull requests that appear to be unreviewed AI output will be closed without review.
+
+### Requirements when AI is used
+
+If AI is used to generate any portion of the code, contributors must adhere to the following requirements:
+
+1. **Explicitly disclose the manner in which AI was employed.** The PR template asks for this. Be honest, this won't automatically disqualify your PR. We'd rather have an honest disclosure than find out later. Trust matters more than method.
+2. **Perform a comprehensive manual review prior to submitting the pull request.** Don't submit code you haven't read carefully and tested locally.
+3. **Be prepared to explain every line of code you submitted when asked about it by a maintainer.** If you can't explain why something works the way it does, you're not ready to submit it.
+4. **Check for an existing pull request addressing the same change.** If one exists, comment there and work with its author instead of opening a duplicate.
+5. **It is strictly prohibited to use AI to write your posts for you** (bug reports, feature requests, pull request descriptions, GitHub discussions, responding to humans, etc.). We need to hear from _you_, not your AI assistant. These are the spaces where we build trust and understanding with contributors, and that only works if we're talking to each other.
+
+### Established contributors
+
+Contributors with a long history of thoughtful, quality contributions to Frigate have earned trust through that track record. The level of scrutiny we apply to AI usage naturally reflects that trust. This isn't a formal exemption, it's just how trust works. If you've been around, we know how you think and how you work. If you're new, we're still getting to know you, and clear disclosure helps build that relationship.
+
+### What this means in practice
+
+We're not trying to gatekeep how you write code. Use whatever tools make you productive. But there's a difference between using AI as a tool to implement something you understand and handing a feature request to an AI and submitting whatever comes back. The former is fine. The latter creates maintenance risk for the project.
+
+Some honest context: when we review a PR, we're not just evaluating whether the code works today. We're evaluating whether we can maintain it, debug it, and extend it long-term, often without the original author's involvement. Code that the author doesn't deeply understand is code that nobody understands, and that's a liability.
+
+One more thing worth saying directly: most maintainers already have access to the same AI tools you do. A PR that's entirely AI-generated, where the author can't explain the design, debug issues independently, or engage substantively in design discussions, doesn't offer something we couldn't produce ourselves. What makes a contribution genuinely valuable is the human judgment and domain understanding behind it, as well as the engagement during review that shapes it into something we can confidently take on long-term.
+
+## Our use of AI
+
+The Frigate documentation site has an "Ask AI" search that answers questions from the docs, and we may use AI tooling to help with triage and project management. Like any automated tooling, it is not always right.
+
+If an AI tool leaves a comment on your contribution, treat it the way you would any other comment. If you think it is wrong, say so, and a brief explanation is enough. Maintainers always have the final say.
+
+## Enforcement
+
+Contributions and posts that do not follow this policy will be closed. Depending on the situation, maintainers may also:
+
+- Hide or delete comments that appear to be unreviewed AI output
+- Mark automated content as spam
+- Close an issue, discussion, or pull request without further review
+- Lock a conversation
+- Temporarily or permanently block an account from participating in the project
+
+Repeated violations may result in being blocked from contributing to Frigate.
+
+### When we get it wrong
+
+There is no reliable way to detect this, and we're not going to pretend otherwise. Whether something reads as unreviewed AI output is a judgment call, usually made quickly, by a volunteer with limited time and no way to know for certain. These calls are subjective and we won't always get them right.
+
+If it happens to you, just say so. A short reply telling us you wrote it yourself is enough, and we'll take you at your word and pick the conversation back up. We would much rather occasionally reopen something we misjudged than treat everyone who posts here as a suspect.
+
+We'd ask for some understanding in return. These calls get made quickly because the volume is real, and time spent second-guessing them is time not spent helping the person in the next thread.
+
+## Attribution
+
+Portions of this policy are adapted from the [Open Home Foundation AI Policy](https://developers.home-assistant.io/docs/ai_policy/).
@@ -2,6 +2,8 @@

 Thank you for your interest in contributing to Frigate. This document covers the expectations and guidelines for contributions. Please read it before submitting a pull request.

+All participation in this project, including pull requests, issues, and discussions, is covered by our [AI policy](AI_POLICY.md).
+
 ## Before you start

 ### Bugfixes
@@ -21,28 +23,16 @@ Before writing code for a new feature:

 ## AI usage policy

-AI tools are a reality of modern development and we're not opposed to their use. But we need to understand your relationship with the code you're submitting. The more AI was involved, the more important it is that you've genuinely reviewed, tested, and understood what it produced.
+AI tools are a reality of modern development and we're not opposed to their use. But we need to understand your relationship with the code you're submitting, and we need to hear from you rather than from your AI assistant.

-### Requirements when AI is used
+**Read the [AI policy](AI_POLICY.md) before you open a pull request.** It is short, and it applies to everything you post here. The parts that most often catch people out:

-If AI is used to generate any portion of the code, contributors must adhere to the following requirements:
+- A person has to be in the loop. Don't wire a bot or agent up to open pull requests, issues, or discussions on your behalf.
+- Disclose how AI was used. The PR template asks for this. Be honest, it won't automatically disqualify your PR.
+- Review and test everything you submit, and be prepared to explain every line when asked.
+- Don't use AI to write your PR description or your replies to maintainers.

-1. **Explicitly disclose the manner in which AI was employed.** The PR template asks for this. Be honest — this won't automatically disqualify your PR. We'd rather have an honest disclosure than find out later. Trust matters more than method.
-2. **Perform a comprehensive manual review prior to submitting the pull request.** Don't submit code you haven't read carefully and tested locally.
-3. **Be prepared to explain every line of code they submitted when asked about it by a maintainer.** If you can't explain why something works the way it does, you're not ready to submit it.
-4. **It is strictly prohibited to use AI to write your posts for you** (bug reports, feature requests, pull request descriptions, GitHub discussions, responding to humans, etc.). We need to hear from _you_, not your AI assistant. These are the spaces where we build trust and understanding with contributors, and that only works if we're talking to each other.
-
-### Established contributors
-
-Contributors with a long history of thoughtful, quality contributions to Frigate have earned trust through that track record. The level of scrutiny we apply to AI usage naturally reflects that trust. This isn't a formal exemption — it's just how trust works. If you've been around, we know how you think and how you work. If you're new, we're still getting to know you, and clear disclosure helps build that relationship.
-
-### What this means in practice
-
-We're not trying to gatekeep how you write code. Use whatever tools make you productive. But there's a difference between using AI as a tool to implement something you understand and handing a feature request to an AI and submitting whatever comes back. The former is fine. The latter creates maintenance risk for the project.
-
-Some honest context: when we review a PR, we're not just evaluating whether the code works today. We're evaluating whether we can maintain it, debug it, and extend it long-term — often without the original author's involvement. Code that the author doesn't deeply understand is code that nobody understands, and that's a liability.
-
-One more thing worth saying directly: most maintainers already have access to the same AI tools you do. A PR that's entirely AI-generated — where the author can't explain the design, debug issues independently, or engage substantively in design discussions — doesn't offer something we couldn't produce ourselves. What makes a contribution genuinely valuable is the human judgment and domain understanding behind it, as well as the engagement during review that shapes it into something we can confidently take on long-term.
+Pull requests that appear to be unreviewed AI output will be closed without review.

 ## Pull request guidelines

@@ -1,10 +1,14 @@
 """Convert the default SSDLite MobileNet v2 model to OpenVINO IR.

 Replaces the legacy openvino-dev Model Optimizer conversion. The TensorFlow
-frontend converts the Object Detection API frozen graph natively; the four TF
-outputs are then repacked into the single [1, 1, 100, 7] DetectionOutput-style
-tensor that Frigate's OpenVINO detector expects, and the input is flipped to
-BGR to match the legacy reverse_input_channels behavior.
+frontend translates the Object Detection API pre and post processors literally,
+producing per-class NonMaxSuppression, NonZero ops and map loops with data
+dependent shapes that the GPU plugin handles very badly. Both are cut out the
+way ssd_v2_support.json used to do it: the preprocessor is an identity at the
+native 300x300 input, and the postprocessor becomes a single fused
+DetectionOutput. The result is the [1, 1, 100, 7] tensor that Frigate's
+OpenVINO detector expects, with the input flipped to BGR to match the legacy
+reverse_input_channels behavior.
 """

 import numpy as np
@@ -12,31 +16,91 @@ import openvino as ov
 from openvino import opset8 as ops
 from openvino.preprocess import PrePostProcessor

+MODEL_DIR = "/models/ssdlite_mobilenet_v2_coco_2018_05_09"
+OUTPUT_PATH = "/models/ssdlite_mobilenet_v2.xml"
+INPUT_SHAPE = [1, 300, 300, 3]
+
+# faster_rcnn_box_coder divides the deltas by pipeline.config's y/x/height/width
+# scales of 10/10/5/5, which DetectionOutput expresses as per-prior variances.
+BOX_VARIANCES = np.float32([0.1, 0.1, 0.2, 0.2])
+
 model = ov.convert_model(
-    "/models/ssdlite_mobilenet_v2_coco_2018_05_09/frozen_inference_graph.pb",
-    input=[("image_tensor:0", [1, 300, 300, 3])],
+    f"{MODEL_DIR}/frozen_inference_graph.pb",
+    input=[("image_tensor:0", INPUT_SHAPE)],
 )

-# rows of (image_id, class_id, score, xmin, ymin, xmax, ymax)
-boxes = model.output("detection_boxes:0").get_node().input_value(0)
-classes = model.output("detection_classes:0").get_node().input_value(0)
-scores = model.output("detection_scores:0").get_node().input_value(0)
+nodes = {op.get_friendly_name(): op for op in model.get_ordered_ops()}
+parameter = model.get_parameters()[0]

-# (ymin,xmin,ymax,xmax) -> (xmin,ymin,xmax,ymax)
-boxes = ops.gather(boxes, [1, 0, 3, 2], 2)
-classes = ops.unsqueeze(classes, 2)
-scores = ops.unsqueeze(scores, 2)
-image_id = ops.multiply(scores, np.float32(0.0))
+preprocessor = nodes["Preprocessor/map/TensorArrayStack/TensorArrayGatherV3"]
+box_deltas = nodes["Postprocessor/Reshape_1"].output(0)
+class_scores = nodes["Postprocessor/convert_scores"].output(0)
+anchors_output = nodes["Postprocessor/Reshape"].output(0)

-detections = ops.concat([image_id, classes, scores, boxes], 2)
-detections = ops.unsqueeze(detections, 1)
+# The anchors only depend on the static input shape, so fold them into a
+# constant and drop the generator subgraph with the rest of the postprocessor.
+probe = ov.Core().compile_model(
+    ov.Model([anchors_output, preprocessor.output(0)], [parameter], "probe"), "CPU"
+)
+probe_input = np.random.default_rng(0).integers(0, 255, INPUT_SHAPE, dtype=np.uint8)
+anchors, resized = (out.copy() for out in probe([probe_input]).values())
+
+assert np.allclose(resized, probe_input, atol=1e-3), (
+    "preprocessor is not an identity at 300x300, it cannot be bypassed"
+)
+
+image = ops.convert(parameter, "f32")
+
+for consumer in list(preprocessor.output(0).get_target_inputs()):
+    consumer.replace_source_output(image.output(0))
+
+# (ymin, xmin, ymax, xmax) -> (xmin, ymin, xmax, ymax)
+priors = anchors[:, [1, 0, 3, 2]].astype(np.float32).reshape(-1)
+variances = np.tile(BOX_VARIANCES, len(anchors))
+proposals = ops.constant(np.stack([priors, variances])[np.newaxis])
+
+# (ty, tx, th, tw) -> (dx, dy, dw, dh) for the CENTER_SIZE decode
+box_logits = ops.reshape(ops.gather(box_deltas, [1, 0, 3, 2], 1), [1, -1], False)
+class_preds = ops.reshape(class_scores, [1, -1], False)
+
+detections = ops.detection_output(
+    box_logits,
+    class_preds,
+    proposals,
+    {
+        "background_label_id": 0,
+        "top_k": 100,
+        "keep_top_k": [100],
+        "nms_threshold": 0.6,
+        "confidence_threshold": 0.3,
+        "code_type": "caffe.PriorBoxParameter.CENTER_SIZE",
+        "share_location": True,
+        "variance_encoded_in_target": False,
+        "normalized": True,
+        "clip_before_nms": False,
+        "clip_after_nms": True,
+        "decrease_label_id": False,
+    },
+)
 detections.output(0).get_tensor().set_names({"detection_out"})

-model = ov.Model([detections], model.get_parameters(), "ssdlite_mobilenet_v2")
+model = ov.Model([detections], [parameter], "ssdlite_mobilenet_v2")

 ppp = PrePostProcessor(model)
 ppp.input().tensor().set_layout(ov.Layout("NHWC"))
 ppp.input().preprocess().reverse_channels()
 model = ppp.build()

-ov.save_model(model, "/models/ssdlite_mobilenet_v2.xml", compress_to_fp16=True)
+# Fail the build rather than silently ship the dynamically shaped graph again.
+op_types = [op.get_type_name() for op in model.get_ordered_ops()]
+assert op_types.count("DetectionOutput") == 1, "postprocessor was not fused"
+
+for dynamic_op in ("NonMaxSuppression", "NonZero", "Loop", "TensorIterator"):
+    assert dynamic_op not in op_types, f"{dynamic_op} left in the graph"
+
+output_shape = model.outputs[0].get_partial_shape()
+assert output_shape.is_static and list(output_shape) == [1, 1, 100, 7], (
+    f"unexpected detector output shape {output_shape}"
+)
+
+ov.save_model(model, OUTPUT_PATH, compress_to_fp16=True)
@@ -2,7 +2,7 @@

 set -euxo pipefail

-SQLITE_VEC_VERSION="0.1.3"
+SQLITE_VEC_VERSION="0.1.9"

 source /etc/os-release

@@ -79,7 +79,5 @@ sherpa-onnx==1.12.*
 faster-whisper==1.1.*
 librosa==0.11.*
 soundfile==0.13.*
-# DeGirum detector
-degirum == 0.16.*
 # Memory profiling
 memray == 1.15.*
@@ -1269,78 +1269,3 @@ axengine:
          input_dtype: int
          input_pixel_format: bgr
          labelmap_path: /labelmap/coco-80.txt
-degirumAiServer:
-  title: DeGirum AI Server
-  models:
-    - key: ai-server-inference
-      label: AI Server Inference
-      recommended: true
-      download: |-
-        Launch a DeGirum AI server as a Docker container, then point the detector at it. Add this to your `docker-compose.yml`:
-
-        ```yaml
-        degirum_detector:
-          container_name: degirum
-          image: degirum/aiserver:latest
-          privileged: true
-          ports:
-            - "8778:8778"
-        ```
-
-        Set `location` to the server's service name, container name, or `host:port`.
-      ui: |
-        Navigate to **Settings > System > Detectors and model** and select **DeGirum** from the detector type dropdown and click **Add**.
-
-        | Field | Value |
-        | --- | --- |
-        | **Location** | `degirum` |
-        | **Zoo** | `degirum/public` |
-        | **Token** | your AI Hub token (optional for the public zoo) |
-      yaml: |
-        degirum_detector:
-          type: degirum
-          location: degirum
-          zoo: degirum/public
-          token: dg_example_token
-degirumLocal:
-  title: DeGirum Local
-  models:
-    - key: local-inference
-      label: Local Inference
-      recommended: true
-      download: Run hardware directly inside the Frigate container with `@local`, removing the AI server hop. The matching device runtime (e.g. the Hailo runtime) must be installed in the container; confirm it with `degirum sys-info`.
-      ui: |
-        Navigate to **Settings > System > Detectors and model** and select **DeGirum** from the detector type dropdown and click **Add**.
-
-        | Field | Value |
-        | --- | --- |
-        | **Location** | `@local` |
-        | **Zoo** | `degirum/public` |
-        | **Token** | your AI Hub token (optional for the public zoo) |
-      yaml: |
-        degirum_detector:
-          type: degirum
-          location: @local
-          zoo: degirum/public
-          token: dg_example_token
-degirumCloud:
-  title: DeGirum AI Hub Cloud
-  models:
-    - key: ai-hub-cloud-inference
-      label: AI Hub Cloud Inference
-      recommended: true
-      download: Run inferences on DeGirum's [AI Hub](https://hub.degirum.com) cloud with `@cloud`. Sign up, create an access token, and set it as `token`. Network latency may require lowering your detection fps.
-      ui: |
-        Navigate to **Settings > System > Detectors and model** and select **DeGirum** from the detector type dropdown and click **Add**.
-
-        | Field | Value |
-        | --- | --- |
-        | **Location** | `@cloud` |
-        | **Zoo** | `degirum/public` |
-        | **Token** | your AI Hub token (optional for the public zoo) |
-      yaml: |
-        degirum_detector:
-          type: degirum
-          location: @cloud
-          zoo: degirum/public
-          token: dg_example_token
@@ -11,6 +11,8 @@ It is not recommended to copy this full configuration file. Only specify values

 :::

+Sections marked `# NOTE: Can be overridden at the camera level` can be set globally and then adjusted per camera. See [Global and Camera-Level Configuration](../config_overrides.md) for how that works.
+
 ```yaml
 mqtt:
  # Optional: Enable mqtt server (default: shown below)
@@ -171,13 +173,14 @@ model:
  # Valid values are rgb, bgr, or yuv. (default: shown below)
  input_pixel_format: rgb
  # Required: Object detection model input tensor format
-  # Valid values are nhwc or nchw (default: shown below)
+  # Valid values are nhwc, nchw, hwnc, or hwcn (default: shown below)
  input_tensor: nhwc
  # Optional: Data type of the model input tensor
  # Valid values are float, float_denorm, or int (default: shown below)
  input_dtype: int
-  # Required: Object detection model type, currently only used with the OpenVINO detector
-  # Valid values are ssd, yolox, yolonas (default: shown below)
+  # Required: Object detection model architecture, used by detectors that support more
+  # than one model type (openvino, onnx, rknn, memryx, axengine, synaptics, and others)
+  # Valid values are ssd, yolox, yolonas, yolo-generic, rfdetr, dfine (default: shown below)
  model_type: ssd
  # Required: Label name modifications. These are merged into the standard labelmap.
  labelmap:
@@ -468,8 +471,8 @@ review:
    detections: False
    # Optional: Activity Context Prompt to give context to the GenAI what activity is and is not suspicious.
    # It is important to be direct and detailed. See documentation for the default prompt structure.
-    activity_context_prompt: """Define what is and is not suspicious
-"""
+    activity_context_prompt: |
+      Define what is and is not suspicious
    # Optional: Image source for GenAI (default: preview)
    # Options: "preview" (uses cached preview frames at ~180p) or "recordings" (extracts frames from recordings at 480p)
    # Using "recordings" provides better image quality but uses more tokens per image.
@@ -813,7 +816,8 @@ classification:
        cameras:
          camera_name:
            # Required: Crop of image frame on this camera to run classification on
-            crop: [0, 180, 220, 400]
+            # [x1, y1, x2, y2] as decimals between 0 and 1, relative to the detect resolution
+            crop: [0.0, 0.25, 0.3, 0.85]
        # Optional: If classification should be run when motion is detected in the crop (default: shown below)
        motion: False
        # Optional: Interval to run classification on in seconds (default: shown below)
@@ -977,7 +981,9 @@ cameras:
      # Optional: Adjust sort order of cameras in the UI. Larger numbers come later (default: shown below)
      # By default the cameras are sorted alphabetically.
      order: 0
-      # Optional: Whether or not to show the camera in the Frigate UI (default: shown below)
+      # Optional: Whether or not to show the camera on the default All Cameras live dashboard.
+      # The camera is still available everywhere else, including camera groups and settings
+      # (default: shown below)
      dashboard: True
      # Optional: Whether this camera is visible in review (the review page and its camera
      # filter, motion review, and the history view) (default: shown below)
@@ -293,6 +293,10 @@ networking:

 This setting is for advanced users. For the majority of use cases it's recommended to change the `ports` section of your Docker compose file or use the Docker `run` `--publish` option instead, e.g. `-p 443:8971`. Changing Frigate's ports may break some integrations.

+The internal and external ports must be different port numbers, and Frigate will refuse to start otherwise. Requests arriving on the internal port are treated as authenticated admins, so pointing both at the same port would remove authentication from the external one.
+
+Nginx binds these ports when it starts, so port changes only take effect after Frigate restarts.
+
 :::

 ### Customizing the Nginx configuration
@@ -335,7 +339,7 @@ For example:
 ```
 services:
  frigate:
-    image: blakeblackshear/frigate:latest
+    image: ghcr.io/blakeblackshear/frigate:stable
    environment:
      - FRIGATE_BASE_PATH=/frigate
 ```
@@ -256,7 +256,7 @@ The only field that is valid at the camera level is `enabled`.

 #### Live transcription

-The single camera Live view in the Frigate UI supports live transcription of audio for streams defined with the `audio` role. Use the Enable/Disable Live Audio Transcription button/switch to toggle transcription processing. When speech is heard, the UI will display a black box over the top of the camera stream with text. The MQTT topic `frigate/<camera_name>/audio/transcription` will also be updated in real-time with transcribed text.
+The single camera Live view in the Frigate UI supports live transcription of audio for streams defined with the `audio` role. Use the Enable/Disable Live Audio Transcription button/switch to toggle transcription processing, or toggle it outside of the UI with the [`frigate/<camera_name>/audio_transcription/set`](/integrations/mqtt#frigatecamera_nameaudio_transcriptionset) MQTT topic or the HTTP API. When speech is heard, the UI will display a black box over the top of the camera stream with text. The MQTT topic `frigate/<camera_name>/audio/transcription` will also be updated in real-time with transcribed text.

 Results can be error-prone due to a number of factors, including:

@@ -272,7 +272,7 @@ If you have CUDA hardware, you can experiment with the `large` `whisper` model o

 #### Transcription and translation of `speech` audio events

-Any `speech` events in Explore can be transcribed and/or translated through the Transcribe button in the Tracked Object Details pane.
+Any `speech` events in Explore can be transcribed and/or translated through the Transcribe button (the microphone icon) in the Tracked Object Details pane.

 In order to use transcription and translation for past events, you must enable audio detection and define `speech` as an audio type to listen for. To have `speech` events translated into the language of your choice, set the `language` config parameter with the correct [language code](https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L10).

@@ -294,7 +294,7 @@ Recorded `speech` events will always use a `whisper` model, regardless of the `m

   Because transcription is **serialized (one event at a time)** and speech events can be generated far faster than they can be processed, an auto-transcribe toggle would very quickly create an ever-growing backlog and degrade core functionality. For the amount of engineering and risk involved, it adds **very little practical value** for the majority of deployments, which are often on low-powered, edge hardware.

-   If you hear speech that's actually important and worth saving/indexing for the future, **just press the transcribe button in Explore** on that specific `speech` event - that keeps things explicit, reliable, and under your control.
+   If you hear speech that's actually important and worth saving/indexing for the future, **just press the transcribe button (the microphone icon) in Explore** on that specific `speech` event - that keeps things explicit, reliable, and under your control.

   Other options are being considered for future versions of Frigate to add transcription options that support external `whisper` Docker containers. A single transcription service could then be shared by Frigate and other applications (for example, Home Assistant Voice), and run on more powerful machines when available.

@@ -165,7 +165,7 @@ If available, recommended settings are:

 #### Setup via the Add Camera Wizard

-The Add Camera Wizard is the recommended way to add a standard Reolink camera. Before starting, make sure [HTTP is enabled](https://support.reolink.com/articles/360003452893-How-to-Access-Reolink-Cameras-NVRs-Home-Hub-Locally-via-Web-Browsers/) in the camera's advanced network settings. The wizard uses the camera's HTTP API to determine its resolution and choose the recommended stream type from the table above.
+The [Add Camera Wizard](cameras.md#adding-a-camera-with-the-add-camera-wizard) is the recommended way to add a standard Reolink camera. Before starting, make sure [HTTP is enabled](https://support.reolink.com/articles/360003452893-How-to-Access-Reolink-Cameras-NVRs-Home-Hub-Locally-via-Web-Browsers/) in the camera's advanced network settings. The wizard uses the camera's HTTP API to determine its resolution and choose the recommended stream type from the table above.

 1. Click **Add Camera** in <NavPath path="Settings > Global configuration > Camera management" />.
 2. Choose **Manual selection** as the stream detection method and select **Reolink** as the camera brand.
@@ -7,6 +7,49 @@ import ConfigTabs from "@site/src/components/ConfigTabs";
 import TabItem from "@theme/TabItem";
 import NavPath from "@site/src/components/NavPath";

+## Adding a camera with the Add Camera Wizard
+
+The Add Camera Wizard is the recommended way to add a camera. Click **Add Camera** in <NavPath path="Settings > Global configuration > Camera management" />. The wizard connects to your camera, tests each stream, and writes the camera's configuration for you, including the [go2rtc](go2rtc.md) restream and the live view stream mapping, so a standard setup needs no hand-written YAML.
+
+### Step 1: Name and connection
+
+Enter a name for the camera along with its host or IP address and credentials, then choose how the wizard should find the camera's streams:
+
+- **Probe camera** queries the camera over ONVIF (the ONVIF port is usually 80 or 8080) and asks it for its stream URLs. Some cameras use a separate ONVIF/service account rather than the device admin user, and some require **Use digest authentication** to be enabled.
+- **Manual selection** builds a stream URL from a template for the camera brand you pick (Dahua/Amcrest/EmpireTech, Hikvision/Uniview/Annke, Ubiquiti, Reolink, Axis, TP-Link, or Foscam). Choose **Other** to enter a custom RTSP URL directly. Non-RTSP stream types must be [configured manually](#setting-up-camera-inputs).
+
+The name you enter is lowercased and spaces become underscores. If the result still isn't a valid config key, the wizard generates a safe name and stores what you typed as `friendly_name`.
+
+### Step 2: Probe or snapshot
+
+In probe mode, the wizard reports what the camera returned (manufacturer, model, firmware, profile count, and whether PTZ, presets, and [autotracking](autotracking.md) are supported) along with the RTSP URLs it discovered. Test each candidate to see its resolution, frame rate, and codecs together with a snapshot, then select the one you want to use.
+
+In manual mode, the wizard tests the templated URL and shows the same metadata and snapshot.
+
+If no RTSP URLs are found, the credentials may be wrong or the camera may not support ONVIF. Go back and use manual selection instead.
+
+### Step 3: Stream configuration
+
+Assign [roles](#setting-up-camera-inputs) to the stream, and use **Add Another Stream** to add the camera's other streams, for example a substream for `detect` alongside the main stream for `record`. At least one stream must have the `detect` role before you can continue.
+
+**Reduce connections to camera** routes that input through the go2rtc restream so Frigate and the live view share a single connection to the camera instead of each opening their own. See [restream](restream.md) for more detail.
+
+### Step 4: Validation and testing
+
+Connect each stream to get a live preview, an estimated bandwidth figure, and a list of validation results. The wizard checks for the most common misconfigurations, including:
+
+- A detect resolution that is too high (increased resource usage) or too low for reliable detection, or one it could not probe at all
+- A stream marked `record` whose audio codec is not AAC, or that has no audio at all
+- A stream marked `audio` that carries no audio stream
+- Using a restreamed input for the `record` role
+- Brand-specific issues, such as an RTSP stream on a Reolink camera that should use http-flv, or a Dahua/Hikvision substream selected for `detect`
+
+**Use stream compatibility mode** passes the stream through go2rtc's ffmpeg module. Enable it if a stream fails to load after several attempts. Note that this also prevents [two way talk](/configuration/live#two-way-talk) from being detected for that stream.
+
+**Save New Camera** writes the configuration and starts the camera right away. No restart is required.
+
+Other features, including [hardware acceleration](hardware_acceleration_video.md), [two way talk](/configuration/live#two-way-talk), and audio transcoding, is configured after the camera has been added. For camera model specific quirks, see the [camera specific](camera_specific.md) docs.
+
 ## Setting Up Camera Inputs

 Several inputs can be configured for each camera and the role of each input can be mixed and matched based on your needs. This allows you to use a lower resolution stream for object detection, but create recordings from a higher resolution stream, or vice versa.
@@ -69,7 +112,7 @@ Additional cameras are simply added under the camera configuration section.
 <ConfigTabs>
 <TabItem value="ui">

-Navigate to <NavPath path="Settings > Global configuration > Camera management" /> and use the add camera button to configure each additional camera.
+Navigate to <NavPath path="Settings > Global configuration > Camera management" /> and use the [Add Camera Wizard](#adding-a-camera-with-the-add-camera-wizard) to configure each additional camera.

 </TabItem>
 <TabItem value="yaml">
@@ -20,7 +20,7 @@ Settings are organized into two scopes:
 - **Global configuration**: values under <NavPath path="Settings > Global configuration" /> apply to every camera by default. This is where you set the baseline behavior for object detection, recording, snapshots, motion, and so on.
 - **Camera configuration**: values under <NavPath path="Settings > Camera configuration" /> apply to a single camera. Use the camera selector button at the top of these pages to choose which camera you are editing.

-When a camera-level section is left untouched, the camera simply inherits the global values. Changing a value on a camera page **overrides** the global value for that camera only: the global setting and every other camera are unaffected. This mirrors how the YAML works, where a value set under `cameras.<name>` takes precedence over the same value set at the top level.
+When a camera-level section is left untouched, the camera simply inherits the global values. Changing a value on a camera page **overrides** the global value for that camera only: the global setting and every other camera are unaffected. This mirrors how the YAML works, where a value set under `cameras.<name>` takes precedence over the same value set at the top level. See [Global and Camera-Level Configuration](./config_overrides.md) for the full details, including how lists and maps are handled and which settings must be enabled globally first.

 To undo an override and go back to inheriting from the parent scope, use the reset button at the bottom of the section:

@@ -130,7 +130,8 @@ go2rtc:

 ```yaml
 genai:
-  api_key: "{FRIGATE_GENAI_API_KEY}"
+  my_provider:
+    api_key: "{FRIGATE_GENAI_API_KEY}"
 ```

 ## Common configuration examples
@@ -0,0 +1,244 @@
+---
+id: config_overrides
+title: Global and Camera-Level Configuration
+---
+
+import ConfigTabs from "@site/src/components/ConfigTabs";
+import TabItem from "@theme/TabItem";
+import NavPath from "@site/src/components/NavPath";
+
+Most of Frigate's configuration can be set once for all cameras and then adjusted for individual cameras. The global value acts as the default for every camera, and any camera can override it.
+
+This page explains how that inheritance works. For a tour of the Settings UI itself, see [Frigate Configuration](./config.md).
+
+## The basics
+
+Set a value globally and every camera uses it. Set the same value on a camera and that camera uses its own value instead.
+
+<ConfigTabs>
+<TabItem value="ui">
+
+1. Navigate to <NavPath path="Settings > Global configuration > Object detection" /> and set **Detect FPS** to `5`. Every camera now detects at 5 fps.
+2. Navigate to <NavPath path="Settings > Camera configuration > Object detection" />, select the `driveway` camera, and set **Detect FPS** to `10`.
+
+The `driveway` camera now detects at 10 fps. Every other camera still uses the global value of 5.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+detect:
+  fps: 5 # every camera detects at 5 fps
+
+cameras:
+  front_door:
+    ffmpeg: ...
+  driveway:
+    ffmpeg: ...
+    detect:
+      fps: 10 # except this one
+```
+
+`front_door` inherits `fps: 5`, and `driveway` uses `10`.
+
+</TabItem>
+</ConfigTabs>
+
+## Overrides apply per value, not per section
+
+Overriding one value in a section does not detach the rest of that section. Everything you don't set on the camera still comes from the global configuration.
+
+<ConfigTabs>
+<TabItem value="ui">
+
+If you set a camera's **Motion threshold** but leave **Contour area** alone, only the threshold is overridden. The contour area continues to follow <NavPath path="Settings > Global configuration > Motion detection" />, and changing it there still affects that camera.
+
+Open a section to see which values are overridden: the section header indicates how many fields differ from the global configuration.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+motion:
+  threshold: 30
+  contour_area: 10
+
+cameras:
+  driveway:
+    motion:
+      threshold: 40
+```
+
+The `driveway` camera ends up with `threshold: 40` and `contour_area: 10`. Only the value you wrote was overridden.
+
+</TabItem>
+</ConfigTabs>
+
+## Returning a camera to the global value
+
+<ConfigTabs>
+<TabItem value="ui">
+
+A camera section that has its own values shows an **Overridden** badge. To remove the override and go back to inheriting, use the **Reset to Global** button at the bottom of the section.
+
+</TabItem>
+<TabItem value="yaml">
+
+Frigate treats a camera value as an override because it is written in the config file, not because it differs from the global value. Repeating the global value under a camera still creates an override:
+
+```yaml
+snapshots:
+  enabled: true
+
+cameras:
+  driveway:
+    snapshots:
+      enabled: true # this is an override, even though it matches
+```
+
+If you later change the global `snapshots.enabled` to `false`, `driveway` keeps saving snapshots, because it has its own value. To make a camera follow the global value again, delete the key from the camera rather than setting it to match.
+
+</TabItem>
+</ConfigTabs>
+
+## Lists replace, maps merge
+
+This is the distinction that surprises people most.
+
+**Lists are replaced entirely.** A camera's list does not add to the global list, it takes its place.
+
+<ConfigTabs>
+<TabItem value="ui">
+
+The camera page shows the objects the camera is currently tracking, starting from the global list. Changing that selection under <NavPath path="Settings > Camera configuration > Objects" /> replaces the list for that camera, so make sure every object you want tracked is selected, not just the ones you are adding.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+objects:
+  track:
+    - person
+    - car
+
+cameras:
+  backyard:
+    objects:
+      track:
+        - dog # backyard tracks ONLY dog, not person or car
+```
+
+To track `dog` in addition to the global objects, list all of them on the camera.
+
+</TabItem>
+</ConfigTabs>
+
+An empty list is a valid override, and is the normal way to opt a camera out of something:
+
+```yaml
+review:
+  alerts:
+    labels:
+      - person
+
+cameras:
+  street:
+    review:
+      alerts:
+        labels: [] # this camera never creates alerts
+```
+
+**Maps are merged key by key.** A camera can add an entry without redeclaring the others.
+
+<ConfigTabs>
+<TabItem value="ui">
+
+Adding a filter for one object under <NavPath path="Settings > Camera configuration > Objects" /> does not remove the filters inherited from <NavPath path="Settings > Global configuration > Objects" />. The camera keeps both.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+objects:
+  filters:
+    person:
+      min_area: 5000
+
+cameras:
+  driveway:
+    objects:
+      filters:
+        car:
+          min_area: 10000
+```
+
+The `driveway` camera ends up with both the `car` filter it defined and the `person` filter from the global configuration.
+
+</TabItem>
+</ConfigTabs>
+
+## Which settings can be overridden
+
+Most, but not all. The [full reference config](./advanced/reference.md) is the authoritative source: sections that support camera-level overrides are marked with the comment `# NOTE: Can be overridden at the camera level`. In the UI, a setting can be overridden if it appears under both <NavPath path="Settings > Global configuration" /> and <NavPath path="Settings > Camera configuration" />.
+
+A few things worth knowing beyond that:
+
+- Some sections are **global only** and have no camera-level equivalent, including `go2rtc`, `genai` providers, `classification`, `telemetry`, `camera_groups`, and `ui`.
+- Some sections exist **only at the camera level**, such as `zones` and `onvif`.
+- Some sections are **partially overridable**, meaning a camera accepts only a few of the keys available globally. `face_recognition`, `lpr`, and `audio_transcription` work this way, and the reference config notes which keys apply.
+
+## Enrichments that must be enabled globally first
+
+License plate recognition and face recognition are special: the global setting is not just a default, it is a switch that must be on before any camera can use the feature. Enabling one on a camera while it is disabled globally is a configuration error, and Frigate will refuse to start:
+
+```
+Camera driveway has lpr enabled but lpr is disabled at the global level of the config. You must enable lpr at the global level.
+```
+
+Enable the feature globally, then turn it off on the cameras that don't need it.
+
+<ConfigTabs>
+<TabItem value="ui">
+
+1. Navigate to <NavPath path="Settings > Global configuration > License plate recognition" /> and enable **LPR**.
+2. Navigate to <NavPath path="Settings > Camera configuration > License plate recognition" />, select each camera that should not run LPR, and disable the **Enable LPR** toggle.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+lpr:
+  enabled: true
+
+cameras:
+  driveway:
+    ffmpeg: ... # inherits lpr, enabled
+  backyard:
+    ffmpeg: ...
+    lpr:
+      enabled: false # opted out
+```
+
+</TabItem>
+</ConfigTabs>
+
+:::note
+
+This applies only to `lpr` and `face_recognition`, because the global setting controls whether the supporting background process starts at all. Other features do not work this way. Audio transcription, for example, can be enabled on a single camera without being enabled globally.
+
+:::
+
+## Profiles
+
+[Profiles](./profiles.md) add a further layer on top of everything described above. A profile is a named set of camera overrides that you can switch on and off while Frigate is running, for example to change detection and recording behavior when you leave the house.
+
+Profiles are applied on top of a camera's already-resolved configuration, so a profile value wins over both the camera and the global value while that profile is active. Profiles cover a subset of the camera sections and do not modify your config file.
+
+## Summary
+
+- A camera inherits every value you don't set on it.
+- Overriding one value does not detach the rest of the section.
+- Writing a value on a camera overrides it, even if it matches the global value. Remove it to inherit again.
+- Lists replace the global list. Maps merge into it.
+- An empty list is an override, not an omission.
+- `lpr` and `face_recognition` must be enabled globally before a camera can use them.
@@ -11,7 +11,7 @@ Object classification allows you to train a custom MobileNetV2 classification mo

 :::info

-Training a custom object classification model requires a one-time internet connection to download MobileNetV2 base weights. Once trained, the model runs fully offline. See [Network Requirements](/frigate/network_requirements#one-time-model-downloads) for details.
+Training a custom object classification model requires an internet connection to download MobileNetV2 base weights. By default these weights are not cached in `/config/`, so they are downloaded again after the container is recreated. Once trained, the model runs fully offline. See [Network Requirements](/frigate/network_requirements#one-time-model-downloads) for details.

 :::

@@ -11,7 +11,7 @@ State classification allows you to train a custom MobileNetV2 classification mod

 :::info

-Training a custom state classification model requires a one-time internet connection to download MobileNetV2 base weights. Once trained, the model runs fully offline. See [Network Requirements](/frigate/network_requirements#one-time-model-downloads) for details.
+Training a custom state classification model requires an internet connection to download MobileNetV2 base weights. By default these weights are not cached in `/config/`, so they are downloaded again after the container is recreated. Once trained, the model runs fully offline. See [Network Requirements](/frigate/network_requirements#one-time-model-downloads) for details.

 :::

@@ -73,9 +73,13 @@ classification:
        interval: 10 # also run every N seconds (optional)
        cameras:
          front:
-            crop: [0, 180, 220, 400]
+            # [x1, y1, x2, y2] as decimals between 0 and 1, relative to the
+            # camera's detect resolution
+            crop: [0.0, 0.25, 0.3, 0.85]
 ```

+Crop coordinates are normalized: each value is a fraction of the camera's `detect` width or height, not a pixel value. Drawing the crop in the UI wizard writes these values for you.
+
 An optional config, `save_attempts`, can be set as a key under the model name. This defines the number of classification attempts to save in the Recent Classifications tab. For state classification models, the default is 100.

 </TabItem>
@@ -232,7 +232,21 @@ Once front-facing images are performing well, start choosing slightly off-angle

 Start with the [Usage](#usage) section and re-read the [Model Requirements](#model-requirements) above.

-1. Ensure `person` is being _detected_. A `person` will automatically be scanned by Frigate for a face. Any detected faces will appear in the Recent Recognitions tab in the Frigate UI's Face Library.
+1. Enable debug logs to see exactly what Frigate is doing.
+   - Enable debug logs for face recognition by adding `frigate.data_processing.real_time.face: debug` to your `logger` configuration. Restart Frigate after this change.
+
+     ```yaml
+     logger:
+       default: info
+       logs:
+         # highlight-next-line
+         frigate.data_processing.real_time.face: debug
+     ```
+
+   - These logs report where the pipeline stopped for each `person` object, such as no face being found within the person's bounding box, the detected face being smaller than `min_area`, or a face being recognized but scoring too low.
+   - If you see no face-related messages at all, also add `frigate.embeddings.maintainer: debug` to confirm that the face processor was created at startup and that `person` updates are reaching it.
+
+2. Ensure `person` is being _detected_. A `person` will automatically be scanned by Frigate for a face. Any detected faces will appear in the Recent Recognitions tab in the Frigate UI's Face Library.

   If you are using a Frigate+ or `face` detecting model:
   - Watch the [debug view](/usage/live#the-single-camera-view) to ensure that `face` is being detected along with `person`.
@@ -242,7 +256,7 @@ Start with the [Usage](#usage) section and re-read the [Model Requirements](#mod
   - Check your `detect` stream resolution and ensure it is sufficiently high enough to capture face details on `person` objects.
   - You may need to lower your `detection_threshold` if faces are not being detected.

-2. Any detected faces will then be _recognized_.
+3. Any detected faces will then be _recognized_.
   - Make sure you have trained at least one face per the recommendations above.
   - Adjust `recognition_threshold` settings per the suggestions [above](#advanced-configuration).

@@ -6,12 +6,46 @@ title: Configuring Generative AI
 import ConfigTabs from "@site/src/components/ConfigTabs";
 import TabItem from "@theme/TabItem";
 import NavPath from "@site/src/components/NavPath";
+import FaqItem from "@site/src/components/FaqItem";

 ## Configuration

-A Generative AI provider can be configured in the global config, which will make the Generative AI features available for use. There are currently 4 native providers available to integrate with Frigate. Other providers that support the OpenAI standard API can also be used. See the OpenAI-Compatible section below.
+A Generative AI provider can be configured in the global config, which will make the Generative AI features available for use. There are currently 5 native providers available to integrate with Frigate. Other providers that support the OpenAI standard API can also be used. See the OpenAI-Compatible section below.

-To use Generative AI, you must define a single provider at the global level of your Frigate configuration. If the provider you choose requires an API key, you may either directly paste it in your configuration, or store it in an environment variable prefixed with `FRIGATE_`.
+`genai` is a map of named providers. Each key under `genai` is a name you choose, and its value is that provider's settings:
+
+<ConfigTabs>
+<TabItem value="ui">
+
+1. Navigate to <NavPath path="Settings > Enrichments > Generative AI" />.
+   - Click **Add** and enter a **Provider name**. Any name of letters, numbers, hyphens, and underscores is accepted, but it cannot be changed from the UI after the provider is created.
+   - Set **Provider** to the service you are using (e.g., `ollama`)
+   - Set **Base URL**, **API key**, and **Model** as required by that provider
+   - Set **Roles** to the roles this provider should handle.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+genai:
+  my_provider: # any name you like
+    provider: ollama
+    base_url: http://localhost:11434
+    model: qwen3-vl:4b
+    roles:
+      - descriptions
+      - embeddings
+      - chat
+```
+
+</TabItem>
+</ConfigTabs>
+
+The examples on this page all use `my_provider`, but the name is arbitrary and is only used to reference the provider elsewhere in the config (for example, `semantic_search.model`).
+
+Each provider handles one or more **roles**: `chat`, `descriptions`, and `embeddings`. A provider handles all three by default, and each role may be assigned to exactly one provider. Define a single provider if you want it to do everything, or split the roles across several providers using the `roles` option.
+
+If the provider you choose requires an API key, you may either directly paste it in your configuration, or store it in an environment variable prefixed with `FRIGATE_`.

 ## Local Providers

@@ -25,15 +59,20 @@ Running Generative AI models on CPU is not recommended, as high inference times

 ### Recommended Local Models

-You must use a vision-capable model with Frigate. The following models are recommended for local deployment:
+You must use a vision-capable model with Frigate. The following models are recommended for local deployment of the `descriptions` and `chat` roles:

 | Model      | Notes                                                                                                                                                                |
 | ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `qwen3-vl` | Strong visual and situational understanding, enhanced ability to identify smaller objects and interactions with object.                                              |
-| `qwen3.5`  | Strong situational understanding, but missing DeepStack from qwen3-vl leading to worse performance for identifying objects in people's hand and other small details. |
-| `qwen3.6`  | Strong situational understanding, similar to qwen3-vl                                                                                                                |
+| `qwen3.6`  | Strong situational understanding, but missing DeepStack from qwen3-vl leading to worse performance for identifying objects in people's hand and other small details. |
 | `gemma4`   | Strong situational understanding, sometimes resorts to more vague terms like 'interacts' instead of assigning a specific action.                                     |

+The `embeddings` role needs a different kind of model. Text queries are matched against the stored image embeddings, so the model must be trained to place images and text into the same vector space. A chat or description model will still return vectors when asked, but those vectors are not trained for retrieval and text searches will return poor matches with no error to indicate why.
+
+| Model                | Notes                                                                                                                                                               |
+| -------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `qwen3-vl-embedding` | Multimodal embeddings for [Semantic Search](/configuration/semantic_search#genai-provider). Must be served by llama.cpp started with `--embeddings` and `--mmproj`. |
+
 :::info

 Each model is available in multiple parameter sizes (3b, 4b, 8b, etc.). Larger sizes are more capable of complex tasks and understanding of situations, but requires more memory and computational resources. It is recommended to try multiple models and experiment to see which performs best.
@@ -78,23 +117,26 @@ All llama.cpp native options can be passed through `provider_options`, including
   - Set **Provider** to `llamacpp`
   - Set **Base URL** to your llama.cpp server address (e.g., `http://localhost:8080`)
   - Set **Model** to the name of your model
-   - Under **Provider Options**, set `context_size` to tell Frigate your context size so it can send the appropriate amount of information
+   - Optionally, under **Provider Options**, set `context_size` to override the context size Frigate detects from the server

 </TabItem>
 <TabItem value="yaml">

 ```yaml
 genai:
-  provider: llamacpp
-  base_url: http://localhost:8080
-  model: your-model-name
-  provider_options:
-    context_size: 16000 # Tell Frigate your context size so it can send the appropriate amount of information.
+  my_provider:
+    provider: llamacpp
+    base_url: http://localhost:8080
+    model: your-model-name
+    provider_options:
+      context_size: 16000 # Optional, overrides the context size reported by the server.
 ```

 </TabItem>
 </ConfigTabs>

+Frigate queries the llama.cpp server for the model's context size at startup and logs it along with the other detected capabilities. If `context_size` is set in `provider_options`, that value is always used instead, even when the server reports its own.
+
 ### Ollama

 [Ollama](https://ollama.com/) allows you to self-host large language models and keep everything running locally. It is highly recommended to host this server on a machine with an Nvidia graphics card, or on a Apple silicon Mac for best performance.
@@ -127,13 +169,14 @@ Note that Frigate will not automatically download the model you specify in your

 ```yaml
 genai:
-  provider: ollama
-  base_url: http://localhost:11434
-  model: qwen3-vl:4b
-  provider_options: # other Ollama client options can be defined
-    keep_alive: -1
-    options:
-      num_ctx: 8192 # make sure the context matches other services that are using ollama
+  my_provider:
+    provider: ollama
+    base_url: http://localhost:11434
+    model: qwen3-vl:4b
+    provider_options: # other Ollama client options can be defined
+      keep_alive: -1
+      options:
+        num_ctx: 8192 # make sure the context matches other services that are using ollama
 ```

 </TabItem>
@@ -149,11 +192,12 @@ For OpenAI-compatible servers (such as llama.cpp) that don't expose the configur

 ```yaml
 genai:
-  provider: openai
-  base_url: http://your-llama-server
-  model: your-model-name
-  provider_options:
-    context_size: 8192 # Specify the configured context size
+  my_provider:
+    provider: openai
+    base_url: http://your-llama-server
+    model: your-model-name
+    provider_options:
+      context_size: 8192 # Specify the configured context size
 ```

 This ensures Frigate uses the correct context window size when generating prompts.
@@ -176,10 +220,11 @@ This ensures Frigate uses the correct context window size when generating prompt

 ```yaml
 genai:
-  provider: openai
-  base_url: http://your-server:port
-  api_key: your-api-key # May not be required for local servers
-  model: your-model-name
+  my_provider:
+    provider: openai
+    base_url: http://your-server:port
+    api_key: your-api-key # May not be required for local servers
+    model: your-model-name
 ```

 </TabItem>
@@ -217,19 +262,21 @@ Ollama also supports [cloud models](https://ollama.com/cloud), where model infer

 ```yaml
 genai:
-  provider: ollama
-  base_url: http://localhost:11434
-  model: cloud-model-name
+  my_provider:
+    provider: ollama
+    base_url: http://localhost:11434
+    model: cloud-model-name
 ```

 or when using Ollama Cloud directly

 ```yaml
 genai:
-  provider: ollama
-  base_url: https://ollama.com
-  model: cloud-model-name
-  api_key: your-api-key
+  my_provider:
+    provider: ollama
+    base_url: https://ollama.com
+    model: cloud-model-name
+    api_key: your-api-key
 ```

 </TabItem>
@@ -267,9 +314,10 @@ To start using Gemini, you must first get an API key from [Google AI Studio](htt

 ```yaml
 genai:
-  provider: gemini
-  api_key: "{FRIGATE_GEMINI_API_KEY}"
-  model: gemini-2.5-flash
+  my_provider:
+    provider: gemini
+    api_key: "{FRIGATE_GEMINI_API_KEY}"
+    model: gemini-2.5-flash
 ```

 </TabItem>
@@ -279,12 +327,13 @@ genai:

 To use a different Gemini-compatible API endpoint, set the `provider_options` with the `base_url` key to your provider's API URL. For example:

-```yaml {4,5}
+```yaml {5,6}
 genai:
-  provider: gemini
-  ...
-  provider_options:
-    base_url: https://...
+  my_provider:
+    provider: gemini
+    ...
+    provider_options:
+      base_url: https://...
 ```

 Other HTTP options are available, see the [python-genai documentation](https://github.com/googleapis/python-genai).
@@ -318,9 +367,10 @@ To start using OpenAI, you must first [create an API key](https://platform.opena

 ```yaml
 genai:
-  provider: openai
-  api_key: "{FRIGATE_OPENAI_API_KEY}"
-  model: gpt-4o
+  my_provider:
+    provider: openai
+    api_key: "{FRIGATE_OPENAI_API_KEY}"
+    model: gpt-4o
 ```

 </TabItem>
@@ -336,13 +386,14 @@ To use a different OpenAI-compatible API endpoint, set the `OPENAI_BASE_URL` env

 For OpenAI-compatible servers (such as llama.cpp) that don't expose the configured context size in the API response, you can manually specify the context size in `provider_options`:

-```yaml {5,6}
+```yaml {6,7}
 genai:
-  provider: openai
-  base_url: http://your-llama-server
-  model: your-model-name
-  provider_options:
-    context_size: 8192 # Specify the configured context size
+  my_provider:
+    provider: openai
+    base_url: http://your-llama-server
+    model: your-model-name
+    provider_options:
+      context_size: 8192 # Specify the configured context size
 ```

 This ensures Frigate uses the correct context window size when generating prompts.
@@ -377,11 +428,91 @@ To start using Azure OpenAI, you must first [create a resource](https://learn.mi

 ```yaml
 genai:
-  provider: azure_openai
-  base_url: https://instance.cognitiveservices.azure.com/openai/responses?api-version=2025-04-01-preview
-  model: gpt-5-mini
-  api_key: "{FRIGATE_OPENAI_API_KEY}"
+  my_provider:
+    provider: azure_openai
+    base_url: https://instance.cognitiveservices.azure.com/openai/responses?api-version=2025-04-01-preview
+    model: gpt-5-mini
+    api_key: "{FRIGATE_OPENAI_API_KEY}"
 ```

 </TabItem>
 </ConfigTabs>
+
+## FAQ
+
+<FaqItem id="how-do-i-debug-genai-issues" question="How do I debug GenAI issues?">
+
+Frigate's Generative AI features are configured and enabled separately. [Review descriptions and summaries](/configuration/genai/genai_review) live under `review.genai`, and [object descriptions](/configuration/genai/genai_objects) live under `objects.genai`. Configuring a provider on this page does not enable either feature, and enabling one does not enable the other. Decide which of the two is not working, then work through the steps below.
+
+1. Confirm a provider is available and holds the `descriptions` role.
+   - Review descriptions, review summaries, and object descriptions all use the provider that has the `descriptions` role assigned in <NavPath path="Settings > Enrichments > Generative AI > Roles" /> (`genai.<provider>.roles`).
+   - A provider is contacted the first time one of its roles is actually used. A provider holding the `embeddings` role for semantic search is initialized during startup, while a `descriptions` provider is not initialized until the first description is requested, which may be well after boot.
+   - In <NavPath path="Settings > Enrichments > Generative AI" />, use **Refresh models** next to the model field. It queries the provider for its model list and is a quick way to verify that the base URL, API key, and network path between Frigate and your provider are correct.
+
+2. Confirm the feature you expect is actually enabled.
+   - Object descriptions are disabled by default. Turn on <NavPath path="Settings > Global configuration > Objects > GenAI object config > Enable GenAI" /> (`objects.genai.enabled`), either globally or per camera. This is the most common reason custom prompts appear to be ignored while review summaries are still being generated.
+   - Review descriptions are disabled by default. Turn on <NavPath path="Settings > Global configuration > Review > GenAI config > Enable GenAI descriptions" /> (`review.genai.enabled`). Once enabled, alerts are described by default but detections are not, so a detection-only review item will never get a summary unless **Enable GenAI for detections** (`review.genai.detections`) is also on.
+
+3. If object descriptions are never requested, check the filters that skip generation.
+   - <NavPath path="Settings > Global configuration > Objects > GenAI object config > GenAI objects" /> (`objects.genai.objects`) limits generation to specific labels, and **Required zones** (`objects.genai.required_zones`) requires the object to have entered one of those zones. If either is set and does not match, Frigate skips the request silently.
+   - Thumbnails are only collected while an object is moving. Objects that go stationary early contribute fewer frames.
+   - **Use snapshots** (`objects.genai.use_snapshot`) requires snapshots to be enabled for the camera. If the snapshot cannot be read, Frigate logs `Cannot load snapshot for <id>, file not found` and no description is generated.
+   - **Send on end** (`objects.genai.send_triggers.tracked_object_end`) is on by default. If you have turned it off in favor of **Early GenAI trigger** (`objects.genai.send_triggers.after_significant_updates`), descriptions are only requested once that number of updates is reached.
+
+4. Enable debug logs to see exactly what Frigate is doing. Restart Frigate after this change. The next step also requires a restart, so turn both on at the same time to avoid restarting twice.
+
+   ```yaml
+   logger:
+     default: info
+     logs:
+       # highlight-start
+       frigate.genai: debug
+       frigate.data_processing.post.object_descriptions: debug
+       frigate.data_processing.post.review_descriptions: debug
+       # highlight-end
+   ```
+
+5. Save the exact images and prompts that were sent to your provider.
+   - Turn on **Save thumbnails** for the feature you are debugging (`review.genai.debug_save_thumbnails` or `objects.genai.debug_save_thumbnails`). Both features write to `/media/frigate/clips/genai-requests/`, and these files are admin-only.
+   - Review descriptions write `genai-requests/<review_id>/` containing the numbered frames that were sent, plus `prompt.txt` and `response.txt` with the exact prompt and the raw, unparsed model response.
+   - Review summary reports write `genai-requests/<start_ts>-<end_ts>/prompt.txt` and `response.txt`. No images are involved, since a report summarizes existing review descriptions.
+   - Object descriptions write `genai-requests/<event_id>/` containing the numbered thumbnails. The prompt for object descriptions is not written to a file, it is only visible in the debug logs from step 4.
+   - Look at the saved images before blaming the model. If the object is small, blurry, or out of frame, no prompt will fix the result. For object descriptions, consider turning on **Use snapshots** (`objects.genai.use_snapshot`) to send a higher quality image. For review items, consider setting **Review image source** (`review.genai.image_source`) to `recordings` for 480p frames instead of the lower resolution preview frames.
+
+<ConfigTabs>
+<TabItem value="ui">
+
+For review descriptions, navigate to <NavPath path="Settings > Global configuration > Review" /> and set **GenAI config > Save thumbnails** to on.
+
+For object descriptions, navigate to <NavPath path="Settings > Global configuration > Objects" />, expand **GenAI object config**, and set **Save thumbnails** to on.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+review:
+  genai:
+    enabled: true
+    # highlight-next-line
+    debug_save_thumbnails: true
+
+objects:
+  genai:
+    enabled: true
+    # highlight-next-line
+    debug_save_thumbnails: true
+```
+
+</TabItem>
+</ConfigTabs>
+
+6. Verify the prompt is what you think it is.
+   - Object description prompts are the ones you control directly. A camera-level <NavPath path="Settings > Camera configuration > Objects > GenAI object config > Caption prompt" /> (`objects.genai.prompt`) overrides the global one, and an entry in **Object prompts** (`objects.genai.object_prompts`) for a label overrides both for that label. Only `{label}`, `{sub_label}`, and `{camera}` are substituted.
+   - Review description prompts are built by Frigate and request a structured JSON response, so they are not fully replaceable. The parts you control are <NavPath path="Settings > Global configuration > Review > GenAI config > Activity context prompt" /> (`review.genai.activity_context_prompt`) and **Additional concerns** (`review.genai.additional_concerns`). Keep the activity context prompt general, since overly specific rules will sway the model's threat level scoring.
+
+7. If descriptions are generated but the results are poor or inconsistent, look at the model and the context window.
+   - Empty fields, missing `shortSummary` values, or `Failed to parse review description` errors usually mean the model is not following the requested JSON schema. Smaller models struggle with structured output. Try a larger parameter size or one of the [recommended models](#recommended-local-models).
+   - Frigate calculates how many frames to send from the context size the provider reports. If your server reports a different value than it is actually running with, frames will be truncated or the request will fail. Pin the value by adding `context_size` under <NavPath path="Settings > Enrichments > Generative AI > Provider options" /> (`genai.<provider>.provider_options`), and for Ollama also confirm `options.num_ctx` there matches the context you have configured.
+   - Check **Review Description Speed** and **Object Description Speed** in <NavPath path="System metrics > Enrichments" />. If inference takes tens of seconds, requests will queue behind each other and descriptions will appear to stop. For Ollama, review `OLLAMA_NUM_PARALLEL`, `OLLAMA_MAX_QUEUE`, and `OLLAMA_MAX_LOADED_MODELS` so that concurrent requests from Frigate are handled the way you expect.
+
+</FaqItem>
@@ -52,9 +52,10 @@ You can define custom prompts at the global level and per-object type. To config

 ```yaml
 genai:
-  provider: ollama
-  base_url: http://localhost:11434
-  model: qwen3-vl:8b-instruct
+  my_provider:
+    provider: ollama
+    base_url: http://localhost:11434
+    model: qwen3-vl:8b-instruct

 objects:
  genai:
@@ -112,3 +113,7 @@ Many providers also have a public facing chat interface for their models. Downlo
 - OpenAI - [ChatGPT](https://chatgpt.com)
 - Gemini - [Google AI Studio](https://aistudio.google.com)
 - Ollama - [Open WebUI](https://docs.openwebui.com/)
+
+## Troubleshooting
+
+If descriptions are not being generated, or the generated descriptions are not what you expect, see [How do I debug GenAI issues?](/configuration/genai/genai_config#how-do-i-debug-genai-issues).
@@ -201,3 +201,7 @@ Along with individual review item summaries, Generative AI can also produce a si
 Review reports can be requested via the [API](/integrations/api/generate-review-summary-review-summarize-start-start-ts-end-end-ts-post) by sending a POST request to `/api/review/summarize/start/{start_ts}/end/{end_ts}` with Unix timestamps.

 For Home Assistant users, there is a built-in service (`frigate.review_summarize`) that makes it easy to request review reports as part of automations or scripts. This allows you to automatically generate daily summaries, vacation reports, or custom time period reports based on your specific needs.
+
+## Troubleshooting
+
+If summaries are not being generated, or the generated summaries are not what you expect, see [How do I debug GenAI issues?](/configuration/genai/genai_config#how-do-i-debug-genai-issues).
@@ -15,7 +15,7 @@ Frigate uses the bundled go2rtc to power a number of key features:

 :::tip[Most users no longer need to configure go2rtc by hand]

-The **camera setup wizard** is the recommended way to add cameras. Click **Add Camera** in <NavPath path="Settings > Global configuration > Camera management" />, and the wizard probes your camera and writes its configuration for you, including the go2rtc restream and the live stream mapping, so go2rtc is set up automatically.
+The [**camera setup wizard**](cameras.md#adding-a-camera-with-the-add-camera-wizard) is the recommended way to add cameras. Click **Add Camera** in <NavPath path="Settings > Global configuration > Camera management" />, and the wizard probes your camera and writes its configuration for you, including the go2rtc restream and the live stream mapping, so go2rtc is set up automatically.

 This guide is mainly useful if you are **upgrading from an older version and have existing cameras that don't yet use go2rtc**, or if you want to fine-tune a stream by hand (for example, to transcode a codec your browser can't play). The [go2rtc troubleshooting guide](/troubleshooting/go2rtc) applies regardless of how your cameras were added.

@@ -67,4 +67,6 @@ If your stream won't play, has no audio, uses excessive CPU, or otherwise misbeh

 ## Homekit Configuration

-To add camera streams to Homekit Frigate must be configured in docker to use `host` networking mode. Once that is done, you can use the go2rtc WebUI (accessed via port 1984, which is disabled by default) to export a camera to Homekit. Any changes made will automatically be saved to `/config/go2rtc_homekit.yml`.
+To export camera streams to HomeKit, Frigate must be configured in docker to use `host` networking mode. HomeKit settings are stored in `/config/go2rtc_homekit.yml` rather than in your Frigate config, and are edited through the go2rtc config editor at `http://<frigate_host>:1984/editor.html`. Pairings are saved back to that file automatically.
+
+See the [HomeKit integration docs](/integrations/homekit) for the full setup, including the video and audio requirements HomeKit places on the stream.
@@ -34,7 +34,7 @@ If you are using go2rtc, you should adjust the following settings in your camera

 - Video codec: **H.264** - provides the most compatible video codec with all Live view technologies and browsers. Avoid any kind of "smart codec" or "+" codec like _H.264+_ or _H.265+_. as these non-standard codecs remove keyframes (see below).
 - Audio codec: **AAC** - provides the most compatible audio codec with all Live view technologies and browsers that support audio.
- I-frame interval (sometimes called the keyframe interval, the interframe space, or the GOP length): match your camera's frame rate, or choose "1x" (for interframe space on Reolink cameras). For example, if your stream outputs 20fps, your i-frame interval should be 20 (or 1x on Reolink). Values higher than the frame rate will cause the stream to take longer to begin playback. See [this page](https://gardinal.net/understanding-the-keyframe-interval/) for more on keyframes. For many users this may not be an issue, but it should be noted that a 1x i-frame interval will cause more storage utilization if you are using the stream for the `record` role as well.
+- I-frame interval (sometimes called the keyframe interval, the interframe space, or the GOP length): match your camera's frame rate, or choose "1x" (for interframe space on Reolink cameras). For example, if your stream outputs 20fps, your i-frame interval should be 20 (or 1x on Reolink). Values higher than the frame rate will cause the stream to take longer to begin playback. See [this page](https://web.archive.org/web/20251213190836/https://gardinal.net/understanding-the-keyframe-interval/) for more on keyframes. For many users this may not be an issue, but it should be noted that a 1x i-frame interval will cause more storage utilization if you are using the stream for the `record` role as well.

 The default video and audio codec on your camera may not always be compatible with your browser, which is why setting them to H.264 and AAC is recommended. See the [go2rtc docs](https://github.com/AlexxIT/go2rtc?tab=readme-ov-file#codecs-madness) for codec support information.

@@ -196,7 +196,7 @@ services:

 :::

-See [go2rtc WebRTC docs](https://github.com/AlexxIT/go2rtc/tree/v1.8.3#module-webrtc) for more information about this.
+See [go2rtc WebRTC docs](https://github.com/AlexxIT/go2rtc/tree/v1.9.14#module-webrtc) for more information about this.

 ### Two way talk

@@ -334,7 +334,7 @@ When your browser runs into problems playing back your camera streams, it will l

 - **stalled**
  - What it means: Playback has stalled because the player has fallen too far behind live (extended buffering or no data arriving).
-  - What to try: This is usually indicative of the browser struggling to decode too many high-resolution streams at once. Try selecting a lower-bandwidth stream (substream), reduce the number of live streams open, improve the network connection, or lower the camera resolution. Also check your camera's keyframe (I-frame) interval: shorter intervals make playback start and recover faster. You can also try increasing the timeout value in the UI pane of Frigate's settings.
+  - What to try: This is usually indicative of the browser struggling to decode too many high-resolution streams at once. Try selecting a lower-bandwidth stream (substream), reduce the number of live streams open, improve the network connection, or lower the camera resolution. Also check your camera's keyframe (I-frame) interval: shorter intervals make playback start and recover faster. You can also try increasing the timeout value in <NavPath path="Settings > UI" /> .

  - Possible console messages from the player code:
    - `Buffer time (10 seconds) exceeded, browser may not be playing media correctly.`
@@ -6,6 +6,7 @@ title: Notifications
 import ConfigTabs from "@site/src/components/ConfigTabs";
 import TabItem from "@theme/TabItem";
 import NavPath from "@site/src/components/NavPath";
+import FaqItem from "@site/src/components/FaqItem";

 # Notifications

@@ -85,7 +86,7 @@ cameras:

 ### Registration

-Once notifications are enabled, press the `Register for Notifications` button on all devices that you would like to receive notifications on. This will register the background worker. After this Frigate must be restarted and then notifications will begin to be sent.
+Once notifications are enabled, press the `Register This Device` button on all devices that you would like to receive notifications on. This will register the background worker. After this Frigate must be restarted and then notifications will begin to be sent.

 ## Supported Notifications

@@ -104,3 +105,62 @@ Different platforms handle notifications differently, some settings changes may
 ### Android

 Most Android phones have battery optimization settings. To get reliable Notification delivery the browser (Chrome, Firefox) should have battery optimizations disabled. If Frigate is running as a PWA then the Frigate app should have battery optimizations disabled as well.
+
+## Notifications FAQ
+
+<FaqItem id="how-do-i-debug-notifications-issues" question="How do I debug notifications issues?">
+
+Push notifications involve Frigate, your browser, and your browser vendor's push service, so it helps to work from the server outward.
+
+1. Enable debug logs for the push client by adding `frigate.comms.webpush: debug` to your `logger` configuration. Restart Frigate after this change.
+
+   ```yaml
+   logger:
+     default: info
+     logs:
+       # highlight-next-line
+       frigate.comms.webpush: debug
+   ```
+
+   These logs show exactly where a notification stopped, including:
+   - `Email must be provided for push notifications to be sent` means the global `email` field is empty and nothing will ever be sent.
+   - `Sending test notification` and `Sending push notification for <camera>, review ID <id>` mean Frigate handed the message off to the push service.
+   - `Skipping notification for <camera> - in global cooldown period` (or `camera-specific cooldown period`) means your [cooldown](#configuration) values suppressed it.
+   - `Notifications for <camera> are currently suspended` means notifications were suspended from <NavPath path="Settings > Notifications" /> or MQTT.
+   - `Notification endpoint expired for <user>, received 410` means that device's subscription is no longer valid and it must be re-registered.
+   - `Failed to send notification to <user> :: <status>` means the push service rejected the message. A `401` or `403` usually points at a VAPID or `email` problem, and a `5xx` is a problem on the push service's end.
+   - If you see no messages at all when an alert occurs, the notification was never queued. Confirm an actual **alert** was created (notifications are not sent for detections), and that notifications are enabled both globally and for that camera.
+
+2. Verify the basics that most reports come down to:
+   - Frigate must be reached over `https` with a certificate your device trusts. Browsers silently refuse to register a service worker otherwise, and a self-signed certificate that is not installed as trusted on the device will fail.
+   - On iOS, notifications only work when Frigate has been installed to the Home Screen via **Share > Add to Home Screen** and opened from that icon. Safari and Chrome tabs cannot receive web push on iOS.
+   - Each device must be registered individually, and Frigate must be restarted after registering before anything can be sent, including test notifications.
+   - The Frigate server needs outbound internet access to the browser vendor's push service. See [Network Requirements](/frigate/network_requirements#push-notifications).
+
+3. Test from the UI. Use the `Send a test notification` button in <NavPath path="Settings > Notifications" />. If the log shows `Sending test notification` but nothing arrives on the device, the problem is between the push service and your device rather than in Frigate.
+
+4. Check the browser side on the device that is not receiving notifications:
+   - Confirm the site's notification permission is set to **Allow** in your browser or OS settings, and that a focus/do not disturb mode is not hiding them.
+   - In desktop browsers, open Developer Tools > Application > Service Workers and confirm `notifications-worker.js` is registered and activated. Unregistering it and registering the device again will rebuild a broken subscription.
+   - Check the browser console and your reverse proxy logs for failures loading `/notifications-worker.js` or errors on `/api/notifications/register`.
+
+</FaqItem>
+
+<FaqItem id="why-did-notifications-stop-arriving-after-working-for-a-while" question="Why did notifications stop arriving after working for a while?">
+
+Push subscriptions are issued by the browser vendor and can be revoked, most often after a browser update, after clearing site data, or when a device has been offline for an extended period. When this happens the device still appears registered in Frigate, but the push service rejects the message. The debug logs will show `Notification endpoint expired` with a `404` or `410` status.
+
+Unregister and re-register the affected device from <NavPath path="Settings > Notifications" />, then restart Frigate.
+
+</FaqItem>
+
+<FaqItem id="why-am-i-not-getting-notifications-for-one-specific-camera" question="Why am I not getting notifications for one specific camera?">
+
+Work through these in order:
+
+- Notifications are only sent for **alerts**. If the camera is producing detections instead, adjust the camera's `review > alerts > labels` so the objects you care about are classified as alerts.
+- Confirm notifications are enabled for that camera in <NavPath path="Settings > Camera configuration > Notifications" />.
+- Check the camera's `cooldown` value, and remember that the global cooldown applies across all cameras. A busy camera can consume the global cooldown and suppress a quieter one.
+- If [authentication](/configuration/authentication) is enabled with roles, users only receive notifications for the cameras their role grants access to.
+
+</FaqItem>
@@ -24,7 +24,6 @@ Frigate supports multiple different detectors that work on different types of ha
 - [Coral EdgeTPU](#edge-tpu-detector): The Google Coral EdgeTPU is available in USB, Mini PCIe, and m.2 formats allowing for a wide range of compatibility with devices.
 - [Hailo](#hailo-8): The Hailo8 and Hailo8L AI Acceleration module is available in m.2 format with a HAT for RPi devices, offering a wide range of compatibility with devices.
 - <CommunityBadge /> [MemryX](#memryx-mx3): The MX3 Acceleration module is available in m.2 format, offering broad compatibility across various platforms.
- <CommunityBadge /> [DeGirum](#degirum): Service for using hardware devices in the cloud or locally. Hardware and models provided on the cloud on [their website](https://hub.degirum.com).

 **AMD**

@@ -298,6 +297,14 @@ detectors:

 :::

+### Intel NPU host requirements {#intel-npu-requirements}
+
+The NPU firmware is loaded by the host kernel and is not part of the Frigate image. Everything else the NPU needs is bundled in the container, so host NPU libraries should never be mounted in.
+
+Frigate bundles a specific version of Intel's [linux-npu-driver](https://github.com/intel/linux-npu-driver/releases), and the host firmware must come from that release or a newer one. Firmware older than the bundled driver may fail with `MAPPED_INFERENCE_VERSION is NOT compatible with the ELF`, where `Expected` is the version the firmware supports and `received` is the version the bundled compiler produced. Distributions often package older firmware than the driver Frigate ships, so check the build date on the host with `sudo dmesg | grep -i vpu` and update it there if needed.
+
+Intel NPUs cannot be used under Home Assistant OS, which does not include the NPU firmware.
+
 ### Configuration {#configuration-openvino}

 <ModelConfigDropdown detectorTitle="OpenVINO" models={objectDetectorsModels.openvino.models} />
@@ -755,87 +762,6 @@ Explanation of the parameters:
  - **example**: Specifying `output_name = "frigate-{quant}-{input_basename}-{soc}-v{tk_version}"` could result in a model called `frigate-i8-my_model-rk3588-v2.3.0.rknn`.
 - `config`: Configuration passed to `rknn-toolkit2` for model conversion. For an explanation of all available parameters have a look at section "2.2. Model configuration" of [this manual](https://github.com/MarcA711/rknn-toolkit2/releases/download/v2.3.2/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf).

-## DeGirum
-
-DeGirum is a detector that can use any type of hardware listed on [their website](https://hub.degirum.com). DeGirum can be used with local hardware through a DeGirum AI Server, or through the use of `@local`. You can also connect directly to DeGirum's AI Hub to run inferences. **Please Note:** This detector _cannot_ be used for commercial purposes.
-
-### Configuration {#configuration-degirum}
-
-#### AI Server Inference
-
-Before starting with the config file for this section, you must first launch an AI server. DeGirum has an AI server ready to use as a docker container. Add this to your `docker-compose.yml` to get started:
-
-```yaml
-degirum_detector:
-  container_name: degirum
-  image: degirum/aiserver:latest
-  privileged: true
-  ports:
-    - "8778:8778"
-```
-
-All supported hardware will automatically be found on your AI server host as long as relevant runtimes and drivers are properly installed on your machine. Refer to [DeGirum's docs site](https://docs.degirum.com/pysdk/runtimes-and-drivers) if you have any trouble.
-
-Once completed, configure the detector as follows:
-
-<ModelConfigDropdown detectorTitle="DeGirum" models={objectDetectorsModels.degirumAiServer.models} />
-
-Setting up a model in the `config.yml` is similar to setting up an AI server.
-You can set it to:
-
- A model listed on the [AI Hub](https://hub.degirum.com), given that the correct zoo name is listed in your detector
-  - If this is what you choose to do, the correct model will be downloaded onto your machine before running.
- A local directory acting as a zoo. See DeGirum's docs site [for more information](https://docs.degirum.com/pysdk/user-guide-pysdk/organizing-models#model-zoo-directory-structure).
- A path to some model.json.
-
-```yaml
-model:
-  path: ./mobilenet_v2_ssd_coco--300x300_quant_n2x_orca1_1 # directory to model .json and file
-  width: 300 # width is in the model name as the first number in the "int"x"int" section
-  height: 300 # height is in the model name as the second number in the "int"x"int" section
-  input_pixel_format: rgb/bgr # look at the model.json to figure out which to put here
-```
-
-#### Local Inference
-
-It is also possible to eliminate the need for an AI server and run the hardware directly. The benefit of this approach is that you eliminate any bottlenecks that occur when transferring prediction results from the AI server docker container to the frigate one. However, the method of implementing local inference is different for every device and hardware combination, so it's usually more trouble than it's worth. A general guideline to achieve this would be:
-
-1. Ensuring that the frigate docker container has the runtime you want to use. So for instance, running `@local` for Hailo means making sure the container you're using has the Hailo runtime installed.
-2. To double check the runtime is detected by the DeGirum detector, make sure the `degirum sys-info` command properly shows whatever runtimes you mean to install.
-3. Create a DeGirum detector in your configuration.
-
-<ModelConfigDropdown detectorTitle="DeGirum" models={objectDetectorsModels.degirumLocal.models} />
-
-Once `degirum_detector` is setup, you can choose a model through 'model' section in the `config.yml` file.
-
-```yaml
-model:
-  path: mobilenet_v2_ssd_coco--300x300_quant_n2x_orca1_1
-  width: 300 # width is in the model name as the first number in the "int"x"int" section
-  height: 300 # height is in the model name as the second number in the "int"x"int" section
-  input_pixel_format: rgb/bgr # look at the model.json to figure out which to put here
-```
-
-#### AI Hub Cloud Inference
-
-If you do not possess whatever hardware you want to run, there's also the option to run cloud inferences. Do note that your detection fps might need to be lowered as network latency does significantly slow down this method of detection. For use with Frigate, we highly recommend using a local AI server as described above. To set up cloud inferences,
-
-1. Sign up at [DeGirum's AI Hub](https://hub.degirum.com).
-2. Get an access token.
-3. Create a DeGirum detector in your configuration.
-
-<ModelConfigDropdown detectorTitle="DeGirum" models={objectDetectorsModels.degirumCloud.models} />
-
-Once `degirum_detector` is setup, you can choose a model through 'model' section in the `config.yml` file.
-
-```yaml
-model:
-  path: mobilenet_v2_ssd_coco--300x300_quant_n2x_orca1_1
-  width: 300 # width is in the model name as the first number in the "int"x"int" section
-  height: 300 # height is in the model name as the second number in the "int"x"int" section
-  input_pixel_format: rgb/bgr # look at the model.json to figure out which to put here
-```
-
 ## AXERA

 Hardware accelerated object detection is supported on the following SoCs:
@@ -126,7 +126,7 @@ Only the fields you explicitly set in a profile override are applied. All other

 ## Activating Profiles

-Profiles can be activated and deactivated via the Frigate UI, [MQTT](/integrations/mqtt#frigateprofileset), or the Home Assistant integration.
+Profiles can be activated and deactivated via the Frigate UI, [MQTT](/integrations/mqtt#frigateprofileset), the [HTTP API](../integrations/api/camera-set-camera-camera-name-set-feature-sub-command-put.api.mdx), or the Home Assistant integration.

 In the Frigate UI, open the Settings cog and select **Profiles** from the submenu to see all defined profiles. From there you can activate any profile or deactivate the current one. The active profile is indicated in the UI so you always know which profile is in effect.

@@ -232,6 +232,21 @@ No. Only one profile can be active at a time. Activating a new profile automatic

 When you delete a base zone or mask in the Frigate UI, any profile overrides for that entry are deleted automatically as part of the same operation. If you remove a base entry by editing your config file directly and leave a profile override behind, the config will fail validation at startup until the orphaned override is removed as well.

+### How do I make a YAML profile track no objects at all?
+
+Set the tracked object list explicitly to an empty list in the profile:
+
+```yaml
+cameras:
+  front_door:
+    profiles:
+      home:
+        objects:
+          track: []
+```
+
+Leaving the `objects` section empty (or omitting `track`) does not clear the list. Empty sections set no fields, so the profile inherits the full tracked object list from the base config, including anything set at the global level. The same applies to other lists, such as `audio.listen`.
+
 ### Why are some settings missing when I configure a profile override?

 Fields that require a Frigate restart to take effect cannot be overridden by profiles, since profiles are applied at runtime without restarting. Those fields are hidden when editing a profile override and can only be changed on the base configuration.
@@ -163,8 +163,8 @@ genai:
    model: your-model-name
    roles:
      - embeddings
-      - vision
-      - tools
+      - descriptions
+      - chat

 semantic_search:
  enabled: True
@@ -34,6 +34,12 @@ The following models are downloaded automatically the first time their associate
 | [Custom classification](/configuration/custom_classification/state_classification) (training) | MobileNetV2 ImageNet base weights (via Keras)                              | Google storage       |
 | [Audio transcription](/configuration/advanced/system)                                         | Whisper or Sherpa-ONNX streaming model                                     | HuggingFace / OpenAI |

+:::note
+
+The MobileNetV2 base weights are the one exception to the `/config/model_cache/` rule. They are also the only entry that is not downloaded when the feature is enabled: Frigate fetches them when a training run actually starts.
+
+:::
+
 ### Hardware-Specific Detector Models

 If you are using one of the following hardware detectors and have not provided your own model file, a default model will be downloaded on first startup:
@@ -75,7 +81,7 @@ If your Frigate instance has restricted internet access, you can point model dow
 | `HF_ENDPOINT`                       | `https://huggingface.co`            | Semantic search, Sherpa-ONNX, AXEngine models |
 | `GITHUB_ENDPOINT`                   | `https://github.com`                | Face recognition, LPR, RKNN models            |
 | `GITHUB_RAW_ENDPOINT`               | `https://raw.githubusercontent.com` | Bird classification                           |
-| `TF_KERAS_MOBILENET_V2_WEIGHTS_URL` | Google storage (Keras default)      | Custom classification training                |
+| `TF_KERAS_MOBILENET_V2_WEIGHTS_URL` | Unset (Keras uses its own default)  | Custom classification training                |

 ## Optional Cloud Services

@@ -147,9 +153,23 @@ When running as a Home Assistant App, the go2rtc startup script queries the loca
 To run Frigate in an air-gapped or offline environment:

 1. **Pre-download models**: Start Frigate with internet access once with all desired features enabled. Models will be cached in `/config/model_cache/`.
-2. **Disable version check**: Set `telemetry.version_check: false` in your configuration.
-3. **Block outbound model requests**: Set the `HF_HUB_OFFLINE=1` and `TRANSFORMERS_OFFLINE=1` environment variables to prevent HuggingFace and Transformers from attempting any network requests.
-4. **Avoid cloud features**: Do not configure Frigate+, Generative AI providers that require internet, or cloud MQTT brokers.
-5. **Use local model mirrors**: If limited internet is available, set the `HF_ENDPOINT`, `GITHUB_ENDPOINT`, and `GITHUB_RAW_ENDPOINT` environment variables to point to local mirrors.
+2. **Pre-download the training base weights**: If you plan to train custom classification models, set `TF_KERAS_MOBILENET_V2_WEIGHTS_URL` before training, then run one training job while online. Without this variable the base weights are cached outside `/config/` and are lost whenever the container is recreated, so a later training run will fail offline. If the machine never has internet access, copy the weights in manually as described below.
+3. **Disable version check**: Set `telemetry.version_check: false` in your configuration.
+4. **Block outbound model requests**: Set the `HF_HUB_OFFLINE=1` and `TRANSFORMERS_OFFLINE=1` environment variables to prevent HuggingFace and Transformers from attempting any network requests.
+5. **Avoid cloud features**: Do not configure Frigate+, Generative AI providers that require internet, or cloud MQTT brokers.
+6. **Use local model mirrors**: If limited internet is available, set the `HF_ENDPOINT`, `GITHUB_ENDPOINT`, `GITHUB_RAW_ENDPOINT`, and `TF_KERAS_MOBILENET_V2_WEIGHTS_URL` environment variables to point to local mirrors.

 After these steps, Frigate will operate with no outbound internet connections.
+
+### Manually Copying the Training Base Weights
+
+On a machine with internet access, download the weights:
+
+```bash
+curl -L -o mobilenet_v2_weights.h5 \
+  "https://storage.googleapis.com/tensorflow/keras-applications/mobilenet_v2/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_0.35_224_no_top.h5"
+```
+
+Copy the file into your Frigate config volume as `/config/model_cache/MobileNet/mobilenet_v2_weights.h5`, keeping that exact filename, then set the environment variable `TF_KERAS_MOBILENET_V2_WEIGHTS_URL` in your Docker compose file to the URL above and restart Frigate.
+
+The variable must be set even though the URL is never contacted. If it is unset, Frigate ignores the copied file and asks Keras to download the weights instead.
@@ -144,7 +144,7 @@ At this point you should be able to start Frigate and a basic config will be cre

 ### Step 2: Add a camera

-Click the **Add Camera** button in <NavPath path="Settings > Global configuration > Camera management" /> to use the camera setup wizard to get your first camera added into Frigate.
+Click the **Add Camera** button in <NavPath path="Settings > Global configuration > Camera management" /> to use the camera setup wizard to get your first camera added into Frigate. See [Adding a camera with the Add Camera Wizard](../configuration/cameras.md#adding-a-camera-with-the-add-camera-wizard) for a walkthrough of each step.

 ### Step 3: Configure hardware acceleration (recommended)

@@ -281,7 +281,7 @@ For advanced usecases, this behavior can be changed with the [RTSP URL
 template](#options) option. When set, this string will override the default stream
 address that is derived from the default behavior described above. This option supports
 [jinja2 templates](https://jinja.palletsprojects.com/) and has the `camera` dict
-variables from [Frigate API](../integrations/api)
+variables from [Frigate API](/integrations/api/frigate-http-api)
 available for the template. Note that no Home Assistant state is available to the
 template, only the camera dict from Frigate.

@@ -3,35 +3,100 @@ id: homekit
 title: HomeKit
 ---

-Frigate cameras can be integrated with Apple HomeKit through go2rtc. This allows you to view your camera streams directly in the Apple Home app on your iOS, iPadOS, macOS, and tvOS devices.
+Frigate cameras can be exported to Apple HomeKit through go2rtc. Each exported camera appears as an accessory in the Apple Home app on your iOS, iPadOS, macOS, and tvOS devices.

 ## Overview

-HomeKit integration is handled entirely through go2rtc, which is embedded in Frigate. go2rtc provides the necessary HomeKit Accessory Protocol (HAP) server to expose your cameras to HomeKit.
+Exporting cameras is handled entirely through go2rtc, which is embedded in Frigate. go2rtc provides the necessary HomeKit Accessory Protocol (HAP) server, so your camera is published to HomeKit as an accessory in its own right.

-## Setup
+:::note

-All HomeKit configuration and pairing should be done through the **go2rtc WebUI**.
+This is the opposite of importing a HomeKit camera. go2rtc can also pair with an existing HomeKit camera (Aqara, Eve, Eufy, and similar) and use it as a stream source, which is what the `add` page of the go2rtc WebUI is for. That page discovers HomeKit accessories on your network and will not list your Frigate cameras. It is not used for exporting.

-### Accessing the go2rtc WebUI
-
-The go2rtc WebUI is available at:
-
-```
-http://<frigate_host>:1984
-```
-
-Replace `<frigate_host>` with the IP address or hostname of your Frigate server.
-
-### Pairing Cameras
-
-1. Navigate to the go2rtc WebUI at `http://<frigate_host>:1984`
-2. Use the `add` section to add a new camera to HomeKit
-3. Follow the on-screen instructions to generate pairing codes for your cameras
+:::

 ## Requirements

- Frigate must be accessible on your local network using host network_mode
- Your iOS device must be on the same network as Frigate
- Port 1984 must be accessible for the go2rtc WebUI
- For detailed go2rtc configuration options, refer to the [go2rtc documentation](https://github.com/AlexxIT/go2rtc)
+- Frigate must be running with `network_mode: host` so that HomeKit can discover your cameras over mDNS
+- Your Apple device must be on the same network as Frigate
+- Port 1984 must be accessible so you can reach the go2rtc WebUI
+
+HomeKit also places strict limits on the stream itself. go2rtc passes your stream through without resizing or re-encoding it, so the stream you export must already meet these requirements:
+
+- **Video:** H.264 at 1920x1080, 1280x720, or 320x240
+- **Audio:** Opus, mono, 16 kHz
+
+A camera's full resolution stream usually does not qualify. See [Exporting a compatible stream](#exporting-a-compatible-stream) below.
+
+## Configuration
+
+HomeKit settings are stored in `/config/go2rtc_homekit.yml`. This is a separate file from your Frigate config, because go2rtc needs to write your pairings back to it when you pair a device.
+
+Edit it using the go2rtc config editor, which writes to that file directly:
+
+```
+http://<frigate_host>:1984/editor.html
+```
+
+Replace `<frigate_host>` with the IP address or hostname of your Frigate server. The editor will be empty until you add a HomeKit section, since this file holds only your HomeKit settings and not the rest of your go2rtc config.
+
+:::warning
+
+Do not put the `homekit:` section in the `go2rtc:` section of your Frigate config.
+
+Frigate regenerates that config on every startup, so go2rtc cannot save your pairings to it. Pairing will appear to succeed and then fail after the next restart with `PairVerify with unknown client_id`. If the section exists in both places, your saved pairings are erased on every restart.
+
+:::
+
+Add an entry for each camera you want to export. The key must match the name of a go2rtc stream, and the pin must be 8 digits. This is the number the Home app calls the setup code:
+
+```yaml
+homekit:
+  front_door:
+    name: Front Door
+    pin: "12345678"
+```
+
+If the key does not match a go2rtc stream, go2rtc logs `[homekit] missing stream:` at startup and the camera will not appear in the Home app.
+
+:::note
+
+go2rtc derives each accessory's HomeKit identity from this key, so renaming it later means the camera appears as a new accessory and has to be paired again. Settle on the name before you pair.
+
+:::
+
+Frigate keeps only the `homekit:` section of this file when it starts, so do not store streams or other go2rtc settings in it.
+
+### Exporting a compatible stream
+
+If a camera's stream does not meet the requirements listed above, define a scaled restream in your Frigate config and point HomeKit at that stream instead of the original:
+
+```yaml
+go2rtc:
+  streams:
+    front_door:
+      - rtsp://user:password@192.168.1.50:554/stream
+    front_door_homekit:
+      - "ffmpeg:front_door#video=h264#width=1280#height=720#audio=opus/16000"
+```
+
+```yaml
+# /config/go2rtc_homekit.yml
+homekit:
+  front_door_homekit:
+    name: Front Door
+    pin: "12345678"
+```
+
+Add `#hardware=cuda`, `#hardware=vaapi`, or the appropriate value for your system to transcode using your GPU. Note that NVENC cannot encode H.264 wider than 4096 pixels, so very wide streams must be scaled down as shown above rather than only re-encoded.
+
+## Pairing Cameras
+
+1. Restart Frigate after adding the `homekit:` section
+2. In the Apple Home app, choose **Add Accessory**, then **More options** to enter a code manually
+3. Select your camera and enter the pin you configured as the setup code
+4. Confirm that a `pairings:` list now appears under the camera in `/config/go2rtc_homekit.yml`
+
+Pairings are saved back to that file automatically. If step 4 shows no `pairings:` list, check the Frigate log for `[homekit] can't save`, which means the `homekit:` section is missing from `/config/go2rtc_homekit.yml`.
+
+For detailed go2rtc configuration options, refer to the [go2rtc documentation](https://github.com/AlexxIT/go2rtc).
@@ -16,7 +16,7 @@ MQTT requires a network connection to your broker. This is typically local, but
 ### `frigate/available`

 Designed to be used as an availability topic with Home Assistant. Possible message are:
-"online": published when Frigate is running (on startup)
+"online": published once Frigate is running and has published its initial state. Note that this is published on every connection to the broker, so it is republished if the broker restarts or the connection drops and recovers, without Frigate itself restarting.
 "stopped": published when Frigate is stopped normally
 "offline": published automatically by the MQTT broker if Frigate disconnects unexpectedly (via MQTT Will Message)

@@ -390,6 +390,18 @@ Topic to turn audio detection for a camera on and off. Expected values are `ON`

 Topic with current state of audio detection for a camera. Published values are `ON` and `OFF`.

+### `frigate/<camera_name>/audio_transcription/set`
+
+Topic to turn [live audio transcription](/configuration/audio_detectors#live-transcription) for a camera on and off. Expected values are `ON` and `OFF`. Transcribed text is published to `frigate/<camera_name>/audio/transcription`.
+
+`ON` is ignored unless audio transcription is enabled in the config for the camera. Unlike the other camera toggles, this one is not persisted across Frigate restarts.
+
+**NOTE:** Requires audio detection and transcription to be enabled
+
+### `frigate/<camera_name>/audio_transcription/state`
+
+Topic with current state of live audio transcription for a camera. Published values are `ON` and `OFF`.
+
 ### `frigate/<camera_name>/recordings/set`

 Topic to turn recordings for a camera on and off. Expected values are `ON` and `OFF`. The change is persisted across Frigate restarts (see [Runtime toggle persistence](/configuration/live#runtime-toggle-persistence)).
@@ -0,0 +1,238 @@
+---
+id: common_errors
+title: Common Error Messages
+---
+
+import FaqItem from "@site/src/components/FaqItem";
+
+This page is an index of error messages you might see in Frigate's logs, what each one means, and where to go next. It is organized by the kind of problem, not by which component logged the message.
+
+Two things to know before you start:
+
+- **Many of these messages come from FFmpeg, go2rtc, GPU drivers, or the operating system, not from Frigate itself.** Frigate captures and re-logs their output, so the log level shown in the Frigate UI does not always reflect the original severity.
+- **Wrapped errors put the real cause on the next line.** When Frigate logs a generic message like `Error occurred when attempting to maintain recording cache`, the actual exception is logged immediately after it. When a camera's FFmpeg process exits, Frigate logs `The following ffmpeg logs include the last 100 lines prior to exit` and dumps that camera's FFmpeg output. Always read those lines, they are where the answer usually is.
+
+## Camera connection and streams
+
+<FaqItem id="connection-refused-no-route-to-host-401-404" question="Connection refused / No route to host / 401 Unauthorized / 404 Not Found">
+
+These are FFmpeg errors about reaching the camera (or the go2rtc restream). `Connection refused` and `No route to host` mean nothing is listening at that address or the host is unreachable; `401 Unauthorized` is wrong credentials; `404 Not Found` is a wrong stream path (or a `restream` input pointing at a go2rtc stream name that does not exist). A camera that has hit its concurrent-connection limit can also return `refused` or `401` on a URL that works in VLC.
+
+See [go2rtc troubleshooting](/troubleshooting/go2rtc#1-read-the-go2rtc-logs) for how to isolate the stream.
+
+</FaqItem>
+
+<FaqItem id="no-frames-received-in-20-seconds" question="No frames received from <camera> in 20 seconds. Exiting ffmpeg...">
+
+FFmpeg is running but has stopped delivering video for 20 seconds, so Frigate's camera watchdog restarts it. The stream connected at least once, then went quiet: a camera reboot, a network drop, the camera evicting the connection, or a stalled decoder. If it repeats on a loop, the stream is unstable.
+
+</FaqItem>
+
+<FaqItem id="ffmpeg-process-crashed-unexpectedly" question="Ffmpeg process crashed unexpectedly for <camera>">
+
+The detect FFmpeg process exited on its own. This message is only the notification; the cause is in the 100 FFmpeg log lines Frigate dumps right after it (look for a `Failed to sync surface`, `Connection refused`, codec, or audio error in that block). Related watchdog messages include `<camera> exceeded fps limit`, which means the camera is delivering frames faster than `detect.fps` (usually a camera whose real frame rate differs from what is configured).
+
+</FaqItem>
+
+<FaqItem id="non-monotonically-increasing-dts" question="Non-monotonic DTS / non monotonically increasing dts to muxer / Queue input is backward in time">
+
+These are FFmpeg messages indicating the camera sent packets with out-of-order timestamps, either on the video or the audio stream. Timestamp jitter like this is common with WiFi cameras and restreamed or proxied sources; other causes are a camera "Smart Codec" / H.264+ / H.265+ mode or a camera clock that jumps. A sustained flood of these messages usually precedes the stream stalling and the watchdog restarting FFmpeg.
+
+In most cases, the fix is to improve the network, reduce system resource usage, or switch to non-WiFi cameras. In general, WiFi cameras are [not recommended](https://ipcamtalk.com/threads/multiple-cameras-high-bandwidth.77100/#post-861110).
+
+On the video stream, this can affect recordings: because they are copied without re-encoding, FFmpeg cannot fix the timestamps, and the segment muxer often splits early, producing one-second segments and a cache backlog. See [Recordings: segments are only 1 second long](/troubleshooting/recordings#segments-are-only-1-second-long).
+
+On the audio stream, the messages can come from the output's audio encoding. If the audio stream is the problem, it may help to have go2rtc transcode it by adding `#audio=aac` to the camera's go2rtc stream to produce clean timestamps for everything consuming the restream.
+
+</FaqItem>
+
+<FaqItem id="bad-cseq" question="RTP: PT=xx: bad cseq (packet loss / reordering)">
+
+An FFmpeg message meaning RTP packets arrived out of sequence, which almost always means the stream is using UDP transport. Frigate's RTSP presets force TCP, so seeing this points at a custom `input_args`, `preset-rtsp-udp`, or a go2rtc source that is not using TCP. Switch to TCP unless your camera is [UDP-only](/configuration/camera_specific#udp-only-cameras).
+
+</FaqItem>
+
+<FaqItem id="error-while-decoding-mb-non-existing-pps" question="error while decoding MB / non-existing PPS referenced (corrupt frames)">
+
+FFmpeg decoder messages meaning the received video bitstream was incomplete or damaged. A few of these at every stream start are normal (the decoder connected before the first keyframe) and Frigate discards them. A continuous stream of them means real packet loss, from Wi-Fi or a saturated link, an overloaded camera, or an FFmpeg restart loop caused by another problem. Fix the underlying instability rather than the message.
+
+</FaqItem>
+
+<FaqItem id="could-not-find-codec-parameters" question="Could not find codec parameters for stream ... unspecified size">
+
+An FFmpeg message meaning it probed the stream but never saw enough decodable video to determine the frame size, often because the probe window ended before the first keyframe on a long-GOP stream, or because the stream is not delivering usable video. If it is a Reolink HTTP stream, use `preset-http-reolink`, which raises the probe size for exactly this case.
+
+</FaqItem>
+
+## Recording
+
+<FaqItem id="no-new-recording-segments" question="No new recording segments were created for <camera> in the last 120s">
+
+Frigate's record watchdog is restarting the record FFmpeg process because no valid segment has reached the cache. This means the record stream is not connecting or the segments are being rejected (see the audio-codec entry below).
+
+See [Recordings: the record stream isn't connecting](/troubleshooting/recordings#the-record-stream-isnt-connecting).
+
+</FaqItem>
+
+<FaqItem id="invalid-or-missing-video-stream-in-segment" question="Invalid or missing video stream in segment. Discarding.">
+
+A cached recording segment failed validation (no readable video stream) and was deleted. The most common cause is a segment that was truncated because the record FFmpeg process was killed mid-write, so this often appears alongside, and as a consequence of, the record-stream restarts above. A segment containing only audio triggers it too.
+
+</FaqItem>
+
+<FaqItem id="incompatible-audio-codec" question="Recordings silently fail to save (incompatible audio codec)">
+
+Some camera audio codecs (G.711 variants such as `pcm_alaw` and `pcm_mulaw`) cannot be stored in an MP4 container, so segments never finalize even though live view works.
+
+See [Recordings: incompatible audio codec](/troubleshooting/recordings#incompatible-audio-codec-recordings-silently-fail-to-save) for the FFmpeg preset that transcodes the audio to AAC.
+
+</FaqItem>
+
+<FaqItem id="error-maintaining-recording-cache" question="Error occurred when attempting to maintain recording cache">
+
+A generic wrapper; the real exception is on the next log line. Frequently it is `[Errno 28] No space left on device` or `[Errno 17] File exists` on a network share.
+
+See [Recordings cache warnings and errors](/troubleshooting/recordings#i-see-the-message-error--error-occurred-when-attempting-to-maintain-recording-cache), which covers this message and the common `Errno` cases.
+
+</FaqItem>
+
+## Hardware acceleration
+
+<FaqItem id="failed-to-sync-surface" question="Failed to sync surface / Failed to download frame: -5 / Error while filtering">
+
+A VAAPI/QSV hardware frame-sync failure between FFmpeg and the GPU driver, not a Frigate bug. It usually appears when the detect stream is being scaled or decoded on the GPU.
+
+See [GPU: Failed to download frame: -5](/troubleshooting/gpu#failed-to-download-frame--5), which lists the fixes in order (switch VAAPI/QSV preset, change `LIBVA_DRIVER_NAME`, use an H.264 substream, match detect resolution and fps to the stream).
+
+</FaqItem>
+
+<FaqItem id="no-decoder-surfaces-left" question="No decoder surfaces left / Can't allocate a surface">
+
+Both mean the GPU ran out of decode surfaces: `No decoder surfaces left` is NVIDIA NVDEC, `Can't allocate a surface` is Intel QSV. This is surface-pool exhaustion, typically from too many concurrent hardware-decoded cameras on one GPU (consumer NVIDIA cards have a driver-enforced limit on simultaneous decode sessions). Reduce the number of cameras decoding on that GPU, decode some on the CPU, or move to hardware without the session cap.
+
+</FaqItem>
+
+<FaqItem id="nvidia-container-cli-nvml-error" question="nvidia-container-cli: nvml error: driver not loaded">
+
+This comes from the NVIDIA container runtime while starting the container, not from Frigate, and the container never starts. The NVIDIA driver is not loaded on the host. Confirm `nvidia-smi` works on the host itself (not inside the container) before troubleshooting Frigate. In a VM or LXC, the driver must be available inside the guest. See [Hardware: Nvidia GPU](/configuration/hardware_acceleration_video).
+
+</FaqItem>
+
+## Detectors and models
+
+<FaqItem id="illegal-instruction" question="Illegal instruction (core dumped)">
+
+The process was killed by the CPU for executing an unsupported instruction. There are two distinct causes in Frigate:
+
+- **A Coral EdgeTPU** on a newer kernel with an outdated gasket driver. See [EdgeTPU: Illegal instruction](/troubleshooting/edgetpu#attempting-to-load-tpu-as-pci--fatal-python-error-illegal-instruction).
+- **A CPU without AVX/AVX2**, when enabling semantic search, face recognition, license plate recognition, classification, or audio transcription. These features use libraries compiled with AVX and crash immediately on CPUs that lack it (commonly Intel Celeron/Pentium before the 2020 Tiger Lake generation). See the [CPU requirements](/frigate/planning_setup#cpu).
+
+</FaqItem>
+
+<FaqItem id="onnx-invalidprotobuf" question="ONNX Runtime InvalidProtobuf / failed to load model">
+
+ONNX Runtime could not parse the model file. The file exists but its contents are not a valid ONNX model, usually a corrupted or interrupted download in `model_cache`, or the wrong file pointed at by `model.path`. Delete the cached model file so Frigate re-downloads it, and confirm `model.path` points at an actual `.onnx` model. See [ONNX detector configuration](/configuration/object_detectors#onnx).
+
+</FaqItem>
+
+<FaqItem id="cuda-failure-999-901" question="CUDA failure 999 / CUDA failure 901">
+
+ONNX Runtime CUDA errors. `999` (`cudaErrorUnknown`) is a general, unrecoverable CUDA context failure, usually a driver/runtime version mismatch between the host and the container or a GPU in a bad state. `901` is a CUDA-graph capture error, which points at a custom model whose operations are not capture-safe. For `999`, align the host driver with the container's CUDA version and confirm the GPU is healthy.
+
+</FaqItem>
+
+<FaqItem id="openvino-no-supported-devices" question="Can't get OPTIMIZATION_CAPABILITIES property as no supported devices found">
+
+OpenVINO could not find the configured device (usually `GPU` or `NPU`). Most often the `/dev/dri` render node is not passed into the container, or the wrong render node is mapped when an iGPU and a discrete GPU coexist.
+
+See [GPU: no supported devices found](/troubleshooting/gpu#cant-get-optimization_capabilities-property-as-no-supported-devices-found).
+
+</FaqItem>
+
+## Memory and storage
+
+<FaqItem id="fatal-python-error-bus-error" question="Fatal Python error: Bus error">
+
+Frigate ran out of shared memory (`/dev/shm`). The container's `shm_size` is too small for the number and resolution of your detect streams, or you added cameras after startup without increasing it.
+
+See [Calculating required shm-size](/frigate/installation#calculating-required-shm-size). If you cannot increase `shm_size`, lowering the `SHM_MAX_FRAMES` environment variable reduces how many frames Frigate buffers per camera.
+
+</FaqItem>
+
+<FaqItem id="errno-28-no-space-left" question="[Errno 28] No space left on device">
+
+A filesystem is full: the recordings volume (`/media/frigate`), the cache tmpfs (`/tmp/cache`), or `/dev/shm`. Check which one, and note that inode exhaustion can produce this while `df -h` still shows free space.
+
+See [Recordings: No space left on device](/troubleshooting/recordings#i-see-the-message-error--error-occurred-when-attempting-to-maintain-recording-cache).
+
+</FaqItem>
+
+<FaqItem id="container-exits-with-no-logs" question="The container exits or restarts with no error in the logs">
+
+A silent exit is usually the host or container out-of-memory killer. Because `/dev/shm` and `/tmp/cache` are memory-backed, they count against the container's memory limit, so aggressive shm or cache sizing can trigger it. Give the container more memory, or reduce shm/cache sizing, and check the host's OOM messages (`dmesg`).
+
+</FaqItem>
+
+## Database
+
+<FaqItem id="database-is-locked" question="database is locked">
+
+SQLite could not acquire the write lock. Frigate's timeout already scales with camera count, so under normal local-disk operation this essentially only happens when the database is on a network share (SMB/NFS), where file locking is unreliable, or when two instances point at the same file.
+
+See [Database is locked](/troubleshooting/faqs#error-database-is-locked).
+
+</FaqItem>
+
+<FaqItem id="database-disk-image-is-malformed" question="database disk image is malformed">
+
+The SQLite database file is corrupted, typically after hard power loss, a network-share database, or a filesystem with unsafe write semantics. Frigate does not repair it automatically, but the database can usually be recovered by hand.
+
+**Stop Frigate first**, then work on the database file directly (by default `/config/frigate.db`). Start by checking what is actually wrong:
+
+```bash
+sqlite3 frigate.db "PRAGMA integrity_check;"
+```
+
+If the only problems reported are index-related (lines such as `row 14 missing from index recordings_path` or `non-unique entry in index ...`), rebuilding the indexes is usually enough and is the least destructive fix:
+
+```bash
+sqlite3 frigate.db "REINDEX;"
+```
+
+If the integrity check reports page or byte-level corruption instead (for example `Multiple uses for byte 2706 of page 142272`), dump the readable contents into a new database:
+
+```bash
+# dump what can still be read
+sqlite3 frigate.db .dump > frigate.dump
+
+# keep the corrupt file, then rebuild from the dump
+mv frigate.db frigate.db.bak
+cat frigate.dump | sqlite3 frigate.db
+
+# confirm the rebuilt database is clean, this should print "ok"
+sqlite3 frigate.db "PRAGMA integrity_check;"
+```
+
+Rows stored in the corrupted pages cannot be recovered, so expect to lose some tracked objects, review items, or thumbnails. Recordings themselves are files on disk and are not affected.
+
+As a last resort, stop Frigate, delete `frigate.db`, and restart. Frigate recreates it, but existing recordings lose all of their metadata. If a `backup.db` exists next to your database, Frigate wrote it before the last schema migration and restoring it recovers everything up to that point.
+
+Repeat corruption usually points at the underlying storage: move the database off a network share, and on Raspberry Pi check power delivery and the SD card or SSD.
+
+</FaqItem>
+
+## Startup and web access
+
+<FaqItem id="unable-to-start-frigate-in-safe-mode" question="Unable to start Frigate in safe mode / Starting Frigate in safe mode">
+
+When your config fails validation at startup, Frigate prints the validation errors (with line numbers), then starts in **safe mode**: a minimal configuration with no cameras and MQTT disabled, so the UI stays reachable. In safe mode the only available page is the Config Editor, which shows the validation errors so you can fix them, then save and restart. Note that recording retention and storage cleanup do **not** run while in safe mode, so do not leave a low-disk system sitting in it.
+
+`Unable to start Frigate in safe mode` means even the minimal config failed, which points at an error in your `auth`, `proxy`, or `database` section, or a config file that is not valid YAML at all. Safe mode is not sticky; fix the config and restart and Frigate returns to normal.
+
+</FaqItem>
+
+<FaqItem id="502-bad-gateway" question="502 Bad Gateway / connection refused to 127.0.0.1:5001">
+
+The web server is up but the Frigate backend (port 5001) is not answering yet. By far the most common reason is that the page was loaded during startup: the API binds last, after database migrations (which can take minutes on a large database), model downloads, and process startup, while the web server is already serving. Wait for startup to finish. If it persists, the backend has failed to start, and the reason is earlier in the logs. This also explains a `connection refused to 127.0.0.1:5001` seen while loading `/ws`, because every authenticated request first makes an auth subrequest to that port.
+
+</FaqItem>
@@ -39,7 +39,7 @@ The per-clip variation is typically quite low and is mostly an artifact of keyfr

 Debug Replay lets you re-run Frigate's detection pipeline against a section of recorded video without manually configuring a dummy camera. It automatically extracts the recording, creates a temporary camera with the same detection settings as the original, and loops the clip through the pipeline so you can observe detections in real time.

-The replay camera behaves like a live camera feed rather than History's video player: it loops the clip continuously as Frigate analyzes it and has no playback controls, so you cannot pause, scrub, or step through it frame by frame.
+The replay camera behaves like a live camera feed rather than History's video player: it loops the clip continuously as Frigate analyzes it and has no playback controls, so you cannot pause, scrub, or step through it frame by frame. The Debug Replay camera does not save recordings or snapshots or surface anything in Explore, but it otherwise behaves like a regular camera, including running enrichments such as Face Recognition, LPR, and custom classification.

 Debug Replay isn't intended to be a one-stop pane for all Frigate diagnostics or a comprehensive debugging environment for every Frigate feature. It merely makes it easier to spin up a "dummy camera" and perform some common adjustments in real time. You'll still need to use the normal tools (logs, an MQTT client, etc) to debug your feature.

@@ -428,3 +428,19 @@ You'll want to:
 - [Tune your motion detection settings](/configuration/motion_detection) either by editing your config file or by using the UI's Motion Tuner.

 </FaqItem>
+
+<FaqItem id="my-timeline-previews-are-black-after-restarting-frigate-or-recreating-the-container" question="My timeline previews are black after restarting Frigate or recreating the container. Why?">
+
+The scrubbing previews (the timelapse clips shown when dragging the History timeline, the secondary-camera previews, and the preview that plays when hovering a review card) are not recorded continuously. Frigate caches low-resolution preview frames in `/tmp/cache` throughout each hour and only assembles them into a finished preview clip **at the top of the hour**.
+
+In the recommended configuration, `/tmp/cache` is a small in-memory (`tmpfs`) area. When Frigate starts, it tries to restore the current hour's cached frames, so a **soft restart from the UI** preserves them. But if you recreate the Docker container or stop Frigate forcibly by any other means partway through an hour, the in-memory cache is discarded, so no preview clip is produced for that partial hour.
+
+This is expected behavior, not a bug:
+
+- Previews for hours that already completed and were written to disk are unaffected.
+- The next full hour after a restart will generate previews normally.
+- This is unrelated to `shm_size`; increasing shared memory does not change it.
+
+To avoid the gap, use the **Restart Frigate** button in the UI's Settings menu rather than recreating the container when possible.
+
+</FaqItem>
@@ -34,7 +34,7 @@ All of your exports live on the **Exports** page, reachable from the main naviga
 - **Rename** it, and
 - **Delete** it: deleting is the only way an export is removed.

-You can also select multiple exports at once to **delete** them in bulk, or to **add them to** (or **remove them from**) a [case](#cases).
+You can also select multiple exports at once to **delete** them in bulk, or to **add them to** (or **remove them from**) a [case](#cases). To download multiple exports as a zip archive, add them to a **case** and use the Download button there.

 ## Cases

@@ -30,6 +30,7 @@ const sidebars: SidebarsConfig = {
    ],
    Configuration: [
      "configuration/config",
+      "configuration/config_overrides",
      {
        type: "category",
        label: "Detectors",
@@ -165,6 +166,7 @@ const sidebars: SidebarsConfig = {
    ],
    Troubleshooting: [
      "troubleshooting/faqs",
+      "troubleshooting/common_errors",
      "troubleshooting/go2rtc",
      "troubleshooting/recordings",
      "troubleshooting/dummy-camera",
@@ -693,6 +693,43 @@ paths:
        **Access:** Admin role required.

        Set a camera feature state. Use camera_name='*' to target all cameras.
+
+        The value to set is sent in the request body as `{"value": "<value>"}`.
+
+        | Feature | Accepted values |
+        | --- | --- |
+        | `enabled` | `ON`, `OFF` |
+        | `detect` | `ON`, `OFF` |
+        | `motion` | `ON`, `OFF` |
+        | `recordings` | `ON`, `OFF` |
+        | `snapshots` | `ON`, `OFF` |
+        | `audio` | `ON`, `OFF` |
+        | `audio_transcription` | `ON`, `OFF` |
+        | `notifications` | `ON`, `OFF` |
+        | `review_alerts` | `ON`, `OFF` |
+        | `review_detections` | `ON`, `OFF` |
+        | `object_descriptions` | `ON`, `OFF` |
+        | `review_descriptions` | `ON`, `OFF` |
+        | `improve_contrast` | `ON`, `OFF` |
+        | `ptz_autotracker` | `ON`, `OFF` |
+        | `birdseye` | `ON`, `OFF` |
+        | `birdseye_mode` | `CONTINUOUS`, `MOTION`, `OBJECTS` |
+        | `motion_contour_area` | integer |
+        | `motion_threshold` | integer |
+        | `motion_mask` | `ON`, `OFF` |
+        | `object_mask` | `ON`, `OFF` |
+        | `zone` | `ON`, `OFF` |
+        | `profile` | a profile name, or `none` to deactivate |
+
+        `motion_mask`, `object_mask`, and `zone` require the `sub_command` path
+        parameter to be set to the name of the mask or zone. All other features
+        reject a sub-command.
+
+        `profile` applies globally rather than per camera, so it requires
+        `camera_name` to be `*`.
+
+        These features map to the equivalent MQTT topics, which document the
+        behavior of each value in more detail.
      operationId: 
        camera_set_camera__camera_name__set__feature___sub_command__put
      parameters:
@@ -746,6 +783,43 @@ paths:
        **Access:** Admin role required.

        Set a camera feature state. Use camera_name='*' to target all cameras.
+
+        The value to set is sent in the request body as `{"value": "<value>"}`.
+
+        | Feature | Accepted values |
+        | --- | --- |
+        | `enabled` | `ON`, `OFF` |
+        | `detect` | `ON`, `OFF` |
+        | `motion` | `ON`, `OFF` |
+        | `recordings` | `ON`, `OFF` |
+        | `snapshots` | `ON`, `OFF` |
+        | `audio` | `ON`, `OFF` |
+        | `audio_transcription` | `ON`, `OFF` |
+        | `notifications` | `ON`, `OFF` |
+        | `review_alerts` | `ON`, `OFF` |
+        | `review_detections` | `ON`, `OFF` |
+        | `object_descriptions` | `ON`, `OFF` |
+        | `review_descriptions` | `ON`, `OFF` |
+        | `improve_contrast` | `ON`, `OFF` |
+        | `ptz_autotracker` | `ON`, `OFF` |
+        | `birdseye` | `ON`, `OFF` |
+        | `birdseye_mode` | `CONTINUOUS`, `MOTION`, `OBJECTS` |
+        | `motion_contour_area` | integer |
+        | `motion_threshold` | integer |
+        | `motion_mask` | `ON`, `OFF` |
+        | `object_mask` | `ON`, `OFF` |
+        | `zone` | `ON`, `OFF` |
+        | `profile` | a profile name, or `none` to deactivate |
+
+        `motion_mask`, `object_mask`, and `zone` require the `sub_command` path
+        parameter to be set to the name of the mask or zone. All other features
+        reject a sub-command.
+
+        `profile` applies globally rather than per camera, so it requires
+        `camera_name` to be `*`.
+
+        These features map to the equivalent MQTT topics, which document the
+        behavior of each value in more detail.
      operationId: camera_set_camera__camera_name__set__feature__put
      parameters:
        - name: camera_name
@@ -31,7 +31,10 @@ from frigate.api.auth import (
    get_allowed_cameras_for_filter,
    require_role,
 )
-from frigate.api.config_util import swap_runtime_config
+from frigate.api.config_util import (
+    publish_camera_section_updates,
+    swap_runtime_config,
+)
 from frigate.api.defs.query.app_query_parameters import AppTimelineHourlyQueryParameters
 from frigate.api.defs.request.app_body import (
    AppConfigSetBody,
@@ -963,6 +966,17 @@ def config_set(request: Request, body: AppConfigSetBody):
                            body.update_topic, settings
                        )

+                        # a config/cameras/* topic publishes camera copies, a
+                        # global topic the global object. FrigateConfig.parse
+                        # folds some global sections down into every camera,
+                        # and workers read both objects, so any such section
+                        # needs its camera copies sent alongside the global
+                        # publish above.
+                        if body.update_topic == "config/birdseye":
+                            publish_camera_section_updates(
+                                request.app, config, CameraConfigUpdateEnum.birdseye
+                            )
+
            return JSONResponse(
                content=(
                    {
@@ -31,7 +31,7 @@ from frigate.api.media_auth import (
    deny_response_for_media_uri,
    is_role_restricted,
 )
-from frigate.config import AuthConfig, NetworkingConfig, ProxyConfig
+from frigate.config import AuthConfig, ProxyConfig
 from frigate.const import CONFIG_DIR, JWT_SECRET_ENV_VAR, PASSWORD_HASH_ALGORITHM
 from frigate.models import User

@@ -620,18 +620,18 @@ def resolve_role(
 def auth(request: Request):
    auth_config: AuthConfig = request.app.frigate_config.auth
    proxy_config: ProxyConfig = request.app.frigate_config.proxy
-    networking_config: NetworkingConfig = request.app.frigate_config.networking

    success_response = Response("", status_code=202)

-    # handle case where internal port is a string with ip:port
-    internal_port = networking_config.listen.internal
-    if type(internal_port) is str:
-        internal_port = int(internal_port.split(":")[-1])
-
    # dont require auth if the request is on the internal port
-    # this header is set by Frigate's nginx proxy, so it cant be spoofed
-    if int(request.headers.get("x-server-port", default=0)) == internal_port:
+    # this header is set by Frigate's nginx proxy, so it cant be spoofed.
+    # the port is the boot-time snapshot rather than the live config value:
+    # nginx's listeners are fixed at container start, so an in-memory config
+    # change must never move the port that is trusted here
+    if (
+        int(request.headers.get("x-server-port", default=0))
+        == request.app.auth_internal_port
+    ):
        success_response.headers["remote-user"] = "anonymous"
        success_response.headers["remote-role"] = "admin"
        return success_response
@@ -1328,7 +1328,45 @@ def camera_set(
    body: CameraSetBody,
    sub_command: str | None = None,
 ):
-    """Set a camera feature state. Use camera_name='*' to target all cameras."""
+    """Set a camera feature state. Use camera_name='*' to target all cameras.
+
+    The value to set is sent in the request body as `{"value": "<value>"}`.
+
+    | Feature | Accepted values |
+    | --- | --- |
+    | `enabled` | `ON`, `OFF` |
+    | `detect` | `ON`, `OFF` |
+    | `motion` | `ON`, `OFF` |
+    | `recordings` | `ON`, `OFF` |
+    | `snapshots` | `ON`, `OFF` |
+    | `audio` | `ON`, `OFF` |
+    | `audio_transcription` | `ON`, `OFF` |
+    | `notifications` | `ON`, `OFF` |
+    | `review_alerts` | `ON`, `OFF` |
+    | `review_detections` | `ON`, `OFF` |
+    | `object_descriptions` | `ON`, `OFF` |
+    | `review_descriptions` | `ON`, `OFF` |
+    | `improve_contrast` | `ON`, `OFF` |
+    | `ptz_autotracker` | `ON`, `OFF` |
+    | `birdseye` | `ON`, `OFF` |
+    | `birdseye_mode` | `CONTINUOUS`, `MOTION`, `OBJECTS` |
+    | `motion_contour_area` | integer |
+    | `motion_threshold` | integer |
+    | `motion_mask` | `ON`, `OFF` |
+    | `object_mask` | `ON`, `OFF` |
+    | `zone` | `ON`, `OFF` |
+    | `profile` | a profile name, or `none` to deactivate |
+
+    `motion_mask`, `object_mask`, and `zone` require the `sub_command` path
+    parameter to be set to the name of the mask or zone. All other features
+    reject a sub-command.
+
+    `profile` applies globally rather than per camera, so it requires
+    `camera_name` to be `*`.
+
+    These features map to the equivalent MQTT topics, which document the
+    behavior of each value in more detail.
+    """
    dispatcher = request.app.dispatcher
    frigate_config: FrigateConfig = request.app.frigate_config

@@ -3,6 +3,30 @@
 from fastapi import FastAPI

 from frigate.config import FrigateConfig
+from frigate.config.camera.updater import (
+    CameraConfigUpdateEnum,
+    CameraConfigUpdateTopic,
+)
+
+
+def publish_camera_section_updates(
+    app: FastAPI, config: FrigateConfig, update_type: CameraConfigUpdateEnum
+) -> None:
+    """Broadcast every camera's re-resolved value for a global section.
+
+    Global sections are folded into each camera at parse time and the camera
+    copies are what workers read, so send them rather than leave a worker to
+    guess which cameras were inheriting.
+    """
+    for camera_name, camera_config in config.cameras.items():
+        settings = getattr(camera_config, update_type.name, None)
+
+        if settings is None:
+            continue
+
+        app.config_publisher.publish_update(
+            CameraConfigUpdateTopic(update_type, camera_name), settings
+        )


 def swap_runtime_config(app: FastAPI, config: FrigateConfig) -> None:
@@ -16,6 +40,10 @@ def swap_runtime_config(app: FastAPI, config: FrigateConfig) -> None:
    camera the user turned off would silently come back on.
    """
    app.frigate_config = config
+
+    if app.config_holder is not None:
+        app.config_holder.set(config)
+
    app.genai_manager.update_config(config)

    if app.profile_manager is not None:
@@ -1538,15 +1538,18 @@ async def set_description(
    event.data["description"] = new_description
    event.save()

-    # If semantic search is enabled, update the index
-    if request.app.frigate_config.semantic_search.enabled:
-        context: EmbeddingsContext = request.app.embeddings
+    context: EmbeddingsContext | None = request.app.embeddings
+
+    if context is not None:
        if len(new_description) > 0:
-            context.update_description(
-                event_id,
-                new_description,
-            )
+            # If semantic search is enabled, update the index
+            if request.app.frigate_config.semantic_search.enabled:
+                context.update_description(
+                    event_id,
+                    new_description,
+                )
        else:
+            # embeddings are always cleaned up so they don't outlive their description
            context.db.delete_embeddings_description(event_ids=[event_id])

    response_message = (
@@ -1675,9 +1678,11 @@ async def delete_single_event(event_id: str, request: Request) -> dict:
    event.delete_instance()
    Timeline.delete().where(Timeline.source_id == event_id).execute()

-    # If semantic search is enabled, update the index
-    if request.app.frigate_config.semantic_search.enabled:
-        context: EmbeddingsContext = request.app.embeddings
+    # embeddings are always cleaned up, even when semantic search is disabled,
+    # so that they don't outlive their events
+    context: EmbeddingsContext | None = request.app.embeddings
+
+    if context is not None:
        context.db.delete_embeddings_thumbnail(event_ids=[event_id])
        context.db.delete_embeddings_description(event_ids=[event_id])

@@ -35,6 +35,7 @@ from frigate.comms.event_metadata_updater import (
 )
 from frigate.config import FrigateConfig
 from frigate.config.camera.updater import CameraConfigUpdatePublisher
+from frigate.config.holder import ConfigHolder
 from frigate.config.profile_manager import ProfileManager
 from frigate.debug_replay import DebugReplayManager, debug_replay_auto_stop_watchdog
 from frigate.embeddings import EmbeddingsContext
@@ -74,6 +75,7 @@ def create_fastapi_app(
    dispatcher: Dispatcher | None = None,
    profile_manager: ProfileManager | None = None,
    enforce_default_admin: bool = True,
+    config_holder: ConfigHolder | None = None,
 ):
    logger.info("Starting FastAPI app")
    app = FastAPI(
@@ -150,6 +152,8 @@ def create_fastapi_app(
    app.include_router(debug_replay.router)
    # App Properties
    app.frigate_config = frigate_config
+    # snapshot the port nginx bound at startup, the live config can be swapped
+    app.auth_internal_port = frigate_config.networking.listen.internal_port
    app.genai_manager = GenAIClientManager(frigate_config)
    app.embeddings = embeddings
    app.detected_frames_processor = detected_frames_processor
@@ -162,6 +166,7 @@ def create_fastapi_app(
    app.replay_manager = replay_manager
    app.dispatcher = dispatcher
    app.profile_manager = profile_manager
+    app.config_holder = config_holder

    if frigate_config.auth.enabled:
        secret = get_jwt_secret()
@@ -53,6 +53,7 @@ from frigate.util.file import (
 )
 from frigate.util.image import get_image_from_recording, get_image_quality_params
 from frigate.util.media import get_keyframe_before
+from frigate.util.object import create_empty_regions_grid

 logger = logging.getLogger(__name__)

@@ -1083,7 +1084,21 @@ def clear_region_grid(request: Request, camera_name: str):
            status_code=404,
        )

-    Regions.delete().where(Regions.camera == camera_name).execute()
+    # store an empty grid instead of deleting the row so the grid is
+    # rebuilt from newly tracked objects and not from all past history
+    region = {
+        Regions.camera: camera_name,
+        Regions.grid: create_empty_regions_grid(),
+        Regions.last_update: datetime.now().timestamp(),
+    }
+    (
+        Regions.insert(region)
+        .on_conflict(
+            conflict_target=[Regions.camera],
+            update=region,
+        )
+        .execute()
+    )
    return JSONResponse(
        content={"success": True, "message": "Region grid cleared"},
    )
@@ -182,7 +182,7 @@ async def get_motion_search_status_endpoint(
        )

    job = get_motion_search_job(job_id)
-    if not job:
+    if not job or job.camera != camera_name:
        return JSONResponse(
            content={"success": False, "message": "Job not found"},
            status_code=404,
@@ -253,7 +253,7 @@ async def cancel_motion_search_endpoint(
        )

    job = get_motion_search_job(job_id)
-    if not job:
+    if not job or job.camera != camera_name:
        return JSONResponse(
            content={"success": False, "message": "Job not found"},
            status_code=404,
@@ -30,6 +30,7 @@ from frigate.comms.ws import WebSocketClient
 from frigate.comms.zmq_proxy import ZmqProxy
 from frigate.config.camera.updater import CameraConfigUpdatePublisher
 from frigate.config.config import FrigateConfig
+from frigate.config.holder import ConfigHolder
 from frigate.config.profile_manager import ProfileManager
 from frigate.const import (
    CACHE_DIR,
@@ -121,8 +122,18 @@ class FrigateApp:
        self.ptz_metrics: dict[str, PTZMetrics] = {}
        self.processes: dict[str, int] = {}
        self.embeddings: EmbeddingsContext | None = None
-        self.profile_manager: ProfileManager | None = None
-        self.config = config
+        self.config_holder = ConfigHolder(config)
+
+    @property
+    def config(self) -> FrigateConfig:
+        """The current config, not the one Frigate booted with.
+
+        Read through the holder so the deferred watchdog factories below build
+        a replacement process from the config as it is now. There is no setter
+        on purpose: a plain attribute would let a caller pin this back to a
+        single object and reintroduce the staleness.
+        """
+        return self.config_holder.config

    def ensure_dirs(self) -> None:
        dirs = [
@@ -270,7 +281,7 @@ class FrigateApp:
                10
                * len([c for c in self.config.cameras.values() if c.enabled_in_config]),
            ),
-            load_vec_extension=self.config.semantic_search.enabled,
+            load_vec_extension=True,
        )
        models = [
            Event,
@@ -343,25 +354,6 @@ class FrigateApp:
        )
        self.dispatcher.profile_manager = self.profile_manager

-    def restore_active_profile(self) -> None:
-        """Re-activate the persisted profile after subscribers are connected.
-
-        ZMQ PUB/SUB drops messages with no subscribers, so activation must
-        run after every config_updater subscriber is up.
-        """
-        if self.profile_manager is None:
-            return
-
-        persisted = ProfileManager.load_persisted_profile()
-        if persisted and any(
-            persisted in cam.profiles for cam in self.config.cameras.values()
-        ):
-            logger.info("Restoring persisted profile '%s'", persisted)
-            # runtime overrides are layered on top via restore_runtime_state()
-            self.profile_manager.activate_profile(
-                persisted, clear_runtime_overrides=False
-            )
-
    def start_detectors(self) -> None:
        for name in self.config.cameras.keys():
            try:
@@ -610,6 +602,13 @@ class FrigateApp:
        self.start_detectors()
        self.init_dispatcher()
        self.init_profile_manager()
+
+        # workers get a copy of the config and can miss the broadcast below, so
+        # apply both layers here. must stay after init_profile_manager(), which
+        # snapshots the base config that profile deactivation resets to
+        self.profile_manager.restore_persisted_profile_to_config()
+        self.dispatcher.reapply_runtime_state_to_config()
+
        self.init_embeddings_client()
        self.start_video_output_processor()
        self.start_ptz_autotracker()
@@ -624,8 +623,9 @@ class FrigateApp:
        self.start_record_cleanup()
        self.start_watchdog()

-        # restore persisted runtime overrides on top of config
-        self.restore_active_profile()
+        # publish for the recording/review/embeddings processes, which start
+        # before the config can be corrected, and for the retained MQTT states
+        self.profile_manager.restore_persisted_profile()
        self.dispatcher.restore_runtime_state()

        self.init_auth()
@@ -645,6 +645,7 @@ class FrigateApp:
                    self.replay_manager,
                    self.dispatcher,
                    self.profile_manager,
+                    config_holder=self.config_holder,
                ),
                host="127.0.0.1",
                port=5001,
@@ -77,6 +77,11 @@ class MqttClient(Communicator):
                "ON" if camera.audio.enabled_in_config else "OFF",
                retain=True,
            )
+            self.publish(
+                f"{camera_name}/audio_transcription/state",
+                "ON" if camera.audio_transcription.live_enabled else "OFF",
+                retain=True,
+            )
            self.publish(
                f"{camera_name}/detect/state",
                "ON" if camera.detect.enabled else "OFF",
@@ -258,6 +263,7 @@ class MqttClient(Communicator):
            "snapshots",
            "detect",
            "audio",
+            "audio_transcription",
            "motion",
            "improve_contrast",
            "ptz_autotracker",
@@ -13,8 +13,8 @@ class CameraUiConfig(FrigateBaseModel):
    )
    dashboard: bool = Field(
        default=True,
-        title="Show in UI",
-        description="Toggle whether this camera is visible everywhere in the Frigate UI. Disabling this will require manually editing the config to view this camera in the UI again.",
+        title="Show on Live dashboard",
+        description="Toggle whether this camera is visible on the default All Cameras live dashboard. The camera remains available everywhere else in the UI, including camera groups and settings.",
    )
    review: bool = Field(
        default=True,
@@ -640,7 +640,7 @@ class FrigateConfig(FrigateBaseModel):
        # set notifications state
        self.notifications.enabled_in_config = self.notifications.enabled

-        # validate genai: each role (tools, vision, embeddings) at most once
+        # validate genai: each role (chat, descriptions, embeddings) at most once
        role_to_name: dict[GenAIRoleEnum, str] = {}
        for name, genai_cfg in self.genai.items():
            for role in genai_cfg.roles:
@@ -0,0 +1,34 @@
+"""Shared handle on the config object that is current for this instance."""
+
+from .config import FrigateConfig
+
+__all__ = ["ConfigHolder"]
+
+
+class ConfigHolder:
+    """Indirection for the most recently parsed config.
+
+    /api/config/set re-parses yaml into a brand new FrigateConfig instead of
+    mutating the old one, so any reference captured during startup goes stale
+    the first time a user saves. Anything that has to build something after
+    startup, most importantly the watchdog factories that rebuild a crashed
+    process, must read through a holder rather than close over a config
+    object, or the rebuilt process comes back with the config as it was at
+    boot and silently discards every change made since.
+
+    There is deliberately no setter on the read side: the swap runs in exactly
+    one place (frigate.api.config_util.swap_runtime_config) and everyone else
+    only reads.
+    """
+
+    def __init__(self, config: FrigateConfig) -> None:
+        self._config = config
+
+    @property
+    def config(self) -> FrigateConfig:
+        """The config as of the most recent successful save."""
+        return self._config
+
+    def set(self, config: FrigateConfig) -> None:
+        """Install a freshly parsed config as the current one."""
+        self._config = config
@@ -1,10 +1,18 @@
-from pydantic import Field
+from pydantic import Field, model_validator

 from .base import FrigateBaseModel

 __all__ = ["IPv6Config", "ListenConfig", "NetworkingConfig"]


+def parse_listen_port(value: int | str) -> int:
+    """Return the port number from a bare port or an "address:port" value."""
+    if isinstance(value, str):
+        return int(value.split(":")[-1])
+
+    return value
+
+
 class IPv6Config(FrigateBaseModel):
    enabled: bool = Field(
        default=False,
@@ -25,6 +33,21 @@ class ListenConfig(FrigateBaseModel):
        description="External listening port for Frigate (default 8971).",
    )

+    @property
+    def internal_port(self) -> int:
+        return parse_listen_port(self.internal)
+
+    @property
+    def external_port(self) -> int:
+        return parse_listen_port(self.external)
+
+    @model_validator(mode="after")
+    def validate_distinct_ports(self) -> "ListenConfig":
+        if self.internal_port == self.external_port:
+            raise ValueError("internal and external must listen on different ports")
+
+        return self
+

 class NetworkingConfig(FrigateBaseModel):
    ipv6: IPv6Config = Field(
@@ -169,6 +169,93 @@ class ProfileManager:
                self.config.active_profile = None
                self._persist_active_profile(None)

+    def _validate_profile_name(self, profile_name: str | None) -> str | None:
+        """Return an error message if the name is not a defined profile."""
+        if profile_name is not None and profile_name not in self.config.profiles:
+            return f"Profile '{profile_name}' is not defined in the profiles section"
+
+        return None
+
+    def _apply_to_config(
+        self, profile_name: str | None
+    ) -> tuple[dict[str, set[str]], str | None]:
+        """Reset every camera to base, then apply the named profile on top.
+
+        Returns the changed camera/section pairs, plus an error message if
+        applying the profile failed partway through.
+        """
+        changed: dict[str, set[str]] = {}
+
+        self._reset_to_base(changed)
+
+        if profile_name is not None:
+            err = self._apply_profile_overrides(profile_name, changed)
+            if err:
+                return changed, err
+
+        return changed, None
+
+    def apply_profile_to_config(self, profile_name: str | None) -> str | None:
+        """Apply a profile to the in-memory config, without publishing it.
+
+        Safe to call ahead of activate_profile: both reset to the base config
+        first, so the later call re-derives the same state and still reports
+        every section as changed.
+
+        Returns:
+            None on success, or an error message string on failure.
+        """
+        err = self._validate_profile_name(profile_name)
+
+        if err:
+            return err
+
+        return self._apply_to_config(profile_name)[1]
+
+    def _persisted_profile_to_restore(self) -> str | None:
+        """Return the persisted profile name, if it still applies to a camera."""
+        persisted = self.load_persisted_profile()
+
+        if not persisted or not any(
+            persisted in cam.profiles for cam in self.config.cameras.values()
+        ):
+            return None
+
+        return persisted
+
+    def restore_persisted_profile_to_config(self) -> None:
+        """Restore the persisted profile into the config, without publishing.
+
+        Called before worker processes start, so they are handed a config that
+        already carries the profile rather than relying on the broadcast that
+        restore_persisted_profile() sends later.
+        """
+        persisted = self._persisted_profile_to_restore()
+
+        if persisted is None:
+            return
+
+        err = self.apply_profile_to_config(persisted)
+
+        if err:
+            logger.error("Failed to apply persisted profile '%s': %s", persisted, err)
+
+    def restore_persisted_profile(self) -> None:
+        """Re-activate the persisted profile once subscribers are connected.
+
+        The config already carries the profile; this pass publishes it for the
+        processes that start before the config can be corrected, and for the
+        retained MQTT states.
+        """
+        persisted = self._persisted_profile_to_restore()
+
+        if persisted is None:
+            return
+
+        logger.info("Restoring persisted profile '%s'", persisted)
+        # runtime overrides are layered on top by the dispatcher's replay
+        self.activate_profile(persisted, clear_runtime_overrides=False)
+
    def activate_profile(
        self,
        profile_name: str | None,
@@ -187,23 +274,16 @@ class ProfileManager:
        Returns:
            None on success, or an error message string on failure.
        """
-        if profile_name is not None:
-            if profile_name not in self.config.profiles:
-                return (
-                    f"Profile '{profile_name}' is not defined in the profiles section"
-                )
+        err = self._validate_profile_name(profile_name)
+
+        if err:
+            return err

        # Track which camera/section pairs get changed for ZMQ publishing
-        changed: dict[str, set[str]] = {}
+        changed, err = self._apply_to_config(profile_name)

-        # Reset all cameras to base config
-        self._reset_to_base(changed)
-
-        # Apply new profile overrides if activating
-        if profile_name is not None:
-            err = self._apply_profile_overrides(profile_name, changed)
-            if err:
-                return err
+        if err:
+            return err

        # Publish ZMQ updates only for sections that actually changed
        self._publish_updates(changed)
@@ -1,9 +1,12 @@
+import logging
 import sqlite3
 from typing import Any

 import regex
 from playhouse.sqliteq import SqliteQueueDatabase

+logger = logging.getLogger(__name__)
+
 REGEXP_TIMEOUT_SECONDS = 1.0


@@ -28,8 +31,14 @@ class SqliteVecQueueDatabase(SqliteQueueDatabase):

    def _load_vec_extension(self, conn: sqlite3.Connection) -> None:
        conn.enable_load_extension(True)
-        conn.load_extension(self.sqlite_vec_path)
-        conn.enable_load_extension(False)
+
+        try:
+            conn.load_extension(self.sqlite_vec_path)
+        except conn.OperationalError:
+            logger.error("Unable to load the sqlite-vec extension")
+            self.load_vec_extension = False
+        finally:
+            conn.enable_load_extension(False)

    def _register_regexp(self, conn: sqlite3.Connection) -> None:
        def regexp(expr: str, item: str | None) -> bool:
@@ -44,13 +53,33 @@ class SqliteVecQueueDatabase(SqliteQueueDatabase):

        conn.create_function("REGEXP", 2, regexp)

-    def delete_embeddings_thumbnail(self, event_ids: list[str]) -> None:
+    def _delete_embeddings(self, table: str, event_ids: list[str]) -> None:
+        """Delete embeddings for the given events, if the table exists.
+
+        Embeddings outlive the events they belong to when semantic search is
+        disabled, so deletes are attempted regardless of the current config.
+        """
+        if not event_ids or not self.load_vec_extension:
+            return
+
+        # the embeddings tables are only created once semantic search has run
+        cursor = self.execute_sql(
+            "SELECT name FROM sqlite_master WHERE type = 'table' AND name = ?",
+            (table,),
+        )
+
+        if cursor.fetchone() is None:
+            logger.debug("Skipping %s cleanup, table does not exist", table)
+            return
+
        ids = ",".join(["?" for _ in event_ids])
-        self.execute_sql(f"DELETE FROM vec_thumbnails WHERE id IN ({ids})", event_ids)
+        self.execute_sql(f"DELETE FROM {table} WHERE id IN ({ids})", event_ids)
+
+    def delete_embeddings_thumbnail(self, event_ids: list[str]) -> None:
+        self._delete_embeddings("vec_thumbnails", event_ids)

    def delete_embeddings_description(self, event_ids: list[str]) -> None:
-        ids = ",".join(["?" for _ in event_ids])
-        self.execute_sql(f"DELETE FROM vec_descriptions WHERE id IN ({ids})", event_ids)
+        self._delete_embeddings("vec_descriptions", event_ids)

    def drop_embeddings_tables(self) -> None:
        self.execute_sql("""
@@ -93,7 +93,7 @@ class ModelConfig(BaseModel):
    model_type: ModelTypeEnum = Field(
        default=ModelTypeEnum.ssd,
        title="Object Detection Model Type",
-        description="Detector model architecture type (ssd, yolox, yolonas) used by some detectors for optimization.",
+        description="Detector model architecture type (ssd, yolox, yolonas, yolo-generic, rfdetr, dfine) used by some detectors for optimization.",
    )
    _merged_labelmap: dict[int, str] | None = PrivateAttr()
    _colormap: dict[int, tuple[int, int, int]] = PrivateAttr()
@@ -1,157 +0,0 @@
-import logging
-import queue
-from typing import Literal
-
-import numpy as np
-from pydantic import ConfigDict, Field
-
-from frigate.detectors.detection_api import DetectionApi
-from frigate.detectors.detector_config import BaseDetectorConfig
-
-logger = logging.getLogger(__name__)
-DETECTOR_KEY = "degirum"
-
-
-### DETECTOR CONFIG ###
-class DGDetectorConfig(BaseDetectorConfig):
-    """DeGirum detector for running models via DeGirum cloud or local inference services."""
-
-    model_config = ConfigDict(
-        title="DeGirum",
-    )
-
-    type: Literal[DETECTOR_KEY]
-    location: str = Field(
-        default=None,
-        title="Inference Location",
-        description="Location of the DeGirim inference engine (e.g. '@cloud', '127.0.0.1').",
-    )
-    zoo: str = Field(
-        default=None,
-        title="Model Zoo",
-        description="Path or URL to the DeGirum model zoo.",
-    )
-    token: str = Field(
-        default=None,
-        title="DeGirum Cloud Token",
-        description="Token for DeGirum Cloud access.",
-    )
-
-
-### ACTUAL DETECTOR  ###
-class DGDetector(DetectionApi):
-    type_key = DETECTOR_KEY
-
-    def __init__(self, detector_config: DGDetectorConfig):
-        try:
-            import degirum as dg
-        except ModuleNotFoundError:
-            raise ImportError("Unable to import DeGirum detector.") from None
-
-        self._queue = queue.Queue()
-        self._zoo = dg.connect(
-            detector_config.location, detector_config.zoo, detector_config.token
-        )
-
-        logger.debug(f"Models in zoo: {self._zoo.list_models()}")
-
-        self.dg_model = self._zoo.load_model(
-            detector_config.model.path,
-        )
-
-        # Setting input image format to raw reduces preprocessing time
-        self.dg_model.input_image_format = "RAW"
-
-        # Prioritize the most powerful hardware available
-        self.select_best_device_type()
-        # Frigate handles pre processing as long as these are all set
-        input_shape = self.dg_model.input_shape[0]
-        self.model_height = input_shape[1]
-        self.model_width = input_shape[2]
-
-        # Passing in dummy frame so initial connection latency happens in
-        # init function and not during actual prediction
-        frame = np.zeros(
-            (detector_config.model.width, detector_config.model.height, 3),
-            dtype=np.uint8,
-        )
-        # Pass in frame to overcome first frame latency
-        self.dg_model(frame)
-        self.prediction = self.prediction_generator()
-
-    def select_best_device_type(self):
-        """
-        Helper function that selects fastest hardware available per model runtime
-        """
-        types = self.dg_model.supported_device_types
-
-        device_map = {
-            "OPENVINO": ["GPU", "NPU", "CPU"],
-            "HAILORT": ["HAILO8L", "HAILO8"],
-            "N2X": ["ORCA1", "CPU"],
-            "ONNX": ["VITIS_NPU", "CPU"],
-            "RKNN": ["RK3566", "RK3568", "RK3588"],
-            "TENSORRT": ["DLA", "GPU", "DLA_ONLY"],
-            "TFLITE": ["ARMNN", "EDGETPU", "CPU"],
-        }
-
-        runtime = types[0].split("/")[0]
-        # Just create an array of format {runtime}/{hardware} for every hardware
-        # in the value for appropriate key in device_map
-        self.dg_model.device_type = [
-            f"{runtime}/{hardware}" for hardware in device_map[runtime]
-        ]
-
-    def prediction_generator(self):
-        """
-        Generator for all incoming frames. By using this generator, we don't have to keep
-        reconnecting our websocket on every "predict" call.
-        """
-        logger.debug("Prediction generator was called")
-        with self.dg_model as model:
-            while 1:
-                logger.info(f"q size before calling get: {self._queue.qsize()}")
-                data = self._queue.get(block=True)
-                logger.info(f"q size after calling get: {self._queue.qsize()}")
-                logger.debug(
-                    f"Data we're passing into model predict: {data}, shape of data: {data.shape}"
-                )
-                result = model.predict(data)
-                logger.debug(f"Prediction result: {result}")
-                yield result
-
-    def detect_raw(self, tensor_input):
-        # Reshaping tensor to work with pysdk
-        truncated_input = tensor_input.reshape(tensor_input.shape[1:])
-        logger.debug(f"Detect raw was called for tensor input: {tensor_input}")
-
-        # add tensor_input to input queue
-        self._queue.put(truncated_input)
-        logger.debug(f"Queue size after adding truncated input: {self._queue.qsize()}")
-
-        # define empty detection result
-        detections = np.zeros((20, 6), np.float32)
-        # grab prediction
-        res = next(self.prediction)
-
-        # If we have an empty prediction, return immediately
-        if len(res.results) == 0 or len(res.results[0]) == 0:
-            return detections
-
-        i = 0
-        for result in res.results:
-            if i >= 20:
-                break
-
-            detections[i] = [
-                result["category_id"],
-                float(result["score"]),
-                result["bbox"][1] / self.model_height,
-                result["bbox"][0] / self.model_width,
-                result["bbox"][3] / self.model_height,
-                result["bbox"][2] / self.model_width,
-            ]
-            i += 1
-
-        logger.debug(f"Detections output: {detections}")
-        return detections
@@ -9,6 +9,7 @@ from pydantic import ConfigDict, Field

 from frigate.detectors.detection_api import DetectionApi
 from frigate.detectors.detector_config import BaseDetectorConfig, ModelTypeEnum
+from frigate.util.model import xyxy_to_xywh_for_nms

 try:
    from tflite_runtime.interpreter import Interpreter, load_delegate
@@ -297,7 +298,7 @@ class EdgeTpuTfl(DetectionApi):
            # until after filtering out redundant boxes
            # Shift the logit scores to be non-negative (required by cv2)
            indices = cv2.dnn.NMSBoxes(
-                bboxes=boxes_filtered_decoded,
+                bboxes=xyxy_to_xywh_for_nms(boxes_filtered_decoded),
                scores=max_scores_filtered_shiftedpositive,
                score_threshold=(
                    self.min_logit_value + self.logit_shift_to_positive_values
@@ -17,6 +17,7 @@ from frigate.detectors.detector_config import (
    ModelTypeEnum,
 )
 from frigate.util.file import FileLock
+from frigate.util.model import xyxy_to_xywh_for_nms

 logger = logging.getLogger(__name__)

@@ -581,7 +582,7 @@ class MemryXDetector(DetectionApi):
            # Convert coordinates to integers
            x_min, y_min, x_max, y_max = map(int, [x_min, y_min, x_max, y_max])

-            # Append valid detections [class_id, confidence, x, y, width, height]
+            # Append valid detections [class_id, confidence, x_min, y_min, x_max, y_max]
            detections.append([class_id, confidence, x_min, y_min, x_max, y_max])

        final_detections = np.zeros((20, 6), np.float32)
@@ -595,7 +596,7 @@ class MemryXDetector(DetectionApi):
        detections = np.array(detections, dtype=np.float32)

        # Apply Non-Maximum Suppression (NMS)
-        bboxes = detections[:, 2:6].tolist()  # (x_min, y_min, width, height)
+        bboxes = xyxy_to_xywh_for_nms(detections[:, 2:6])
        scores = detections[:, 1].tolist()  # Confidence scores

        indices = cv2.dnn.NMSBoxes(bboxes, scores, 0.45, 0.5)
@@ -226,12 +226,12 @@ class OvDetector(DetectionApi):

            conf_mask = (image_pred[:, 4] * class_conf.squeeze() >= 0.3).squeeze()
            # Detections ordered as (x1, y1, x2, y2, obj_conf, class_conf, class_pred)
-            detections = np.concatenate(
+            predictions = np.concatenate(
                (image_pred[:, :5], class_conf, class_pred), axis=1
            )
-            detections = detections[conf_mask]
+            predictions = predictions[conf_mask]

-            ordered = detections[detections[:, 5].argsort()[::-1]][:20]
+            ordered = predictions[predictions[:, 5].argsort()[::-1]][:20]

            for i, object_detected in enumerate(ordered):
                detections[i] = self.process_yolo(
@@ -12,7 +12,7 @@ from frigate.const import MODEL_CACHE_DIR, SUPPORTED_RK_SOCS
 from frigate.detectors.detection_api import DetectionApi
 from frigate.detectors.detection_runners import RKNNModelRunner
 from frigate.detectors.detector_config import BaseDetectorConfig, ModelTypeEnum
-from frigate.util.model import post_process_yolo
+from frigate.util.model import post_process_yolo, xyxy_to_xywh_for_nms
 from frigate.util.rknn_converter import auto_convert_model

 logger = logging.getLogger(__name__)
@@ -285,7 +285,7 @@ class Rknn(DetectionApi):

        # run nms
        indices = cv2.dnn.NMSBoxes(
-            bboxes=boxes,
+            bboxes=xyxy_to_xywh_for_nms(boxes),
            scores=scores,
            score_threshold=0.4,
            nms_threshold=0.4,
@@ -366,9 +366,10 @@ class EventCleanup(threading.Thread):
                    logger.debug(f"Deleting {len(chunk)} events from the database")
                    Event.delete().where(Event.id << chunk).execute()

-                    if self.config.semantic_search.enabled:
-                        self.db.delete_embeddings_description(event_ids=chunk)
-                        self.db.delete_embeddings_thumbnail(event_ids=chunk)
-                        logger.debug(f"Deleted {len(ids_to_delete)} embeddings")
+                    # embeddings are always cleaned up, even when semantic search
+                    # is disabled, so that they don't outlive their events
+                    self.db.delete_embeddings_description(event_ids=chunk)
+                    self.db.delete_embeddings_thumbnail(event_ids=chunk)
+                    logger.debug(f"Deleted {len(chunk)} embeddings")

        logger.info("Exiting event cleanup...")
@@ -23,6 +23,7 @@ from frigate.genai.prompts import (
    build_review_summary_prompt,
 )
 from frigate.models import Event
+from frigate.util.builtin import has_non_finite_number

 logger = logging.getLogger(__name__)

@@ -164,6 +165,15 @@ class GenAIClient:
                except json.JSONDecodeError as je:
                    logger.error("Failed to parse review description JSON: %s", je)
                    return None
+
+                # model_construct skips validation, so non-finite numbers that
+                # the validated path would have rejected have to be caught here
+                if has_non_finite_number(raw):
+                    logger.error(
+                        "Discarding review description containing non-finite numbers."
+                    )
+                    return None
+
                # observations and confidence are required on the model; fill an empty default
                # if the response omitted it so attribute access stays safe.
                raw.setdefault("observations", [])
@@ -192,7 +192,7 @@ class LlamaCppClient(GenAIClient):
        logger.info(
            "llama.cpp model '%s' initialized — context: %s, vision: %s, audio: %s, tools: %s, reasoning: %s",
            configured_model,
-            self._context_size or "unknown",
+            self.get_context_size(),
            self._supports_vision,
            self._supports_audio,
            self._supports_tools,
@@ -178,13 +178,10 @@ class OutputProcess(FrigateProcess):
            )

            if update_topic is not None and birdseye_config is not None:
-                previous_global_mode = self.config.birdseye.mode
+                # only the global-only fields are applied here; the per-camera
+                # enabled and mode arrive on config/cameras/<name>/birdseye,
+                # already resolved against yaml by the config parse
                self.config.birdseye = birdseye_config
-
-                for camera_config in self.config.cameras.values():
-                    if camera_config.birdseye.mode == previous_global_mode:
-                        camera_config.birdseye.mode = birdseye_config.mode
-
                logger.debug("Applied dynamic birdseye config update")

            # check if there is an updated config
@@ -159,6 +159,8 @@ class FFMpegConverter(threading.Thread):
                f"duration {self.frame_times[t_idx + 1] - self.frame_times[t_idx]}"
            )

+        Path(self.path).parent.mkdir(parents=True, exist_ok=True)
+
        try:
            p = sp.run(
                self.ffmpeg_cmd.split(" "),
@@ -625,14 +625,18 @@ class OnvifController:
            return

        self.cams[camera_name]["active"] = True
-        self.ptz_metrics[camera_name].motor_stopped.clear()
-        logger.debug(
-            f"{camera_name}: PTZ start time: {self.ptz_metrics[camera_name].frame_time.value}"
-        )
-        self.ptz_metrics[camera_name].start_time.value = self.ptz_metrics[
-            camera_name
-        ].frame_time.value
-        self.ptz_metrics[camera_name].stop_time.value = 0
+
+        # only track start_time for autotracking
+        if self.ptz_metrics[camera_name].autotracker_enabled.value:
+            self.ptz_metrics[camera_name].motor_stopped.clear()
+            logger.debug(
+                f"{camera_name}: PTZ start time: {self.ptz_metrics[camera_name].frame_time.value}"
+            )
+            self.ptz_metrics[camera_name].start_time.value = self.ptz_metrics[
+                camera_name
+            ].frame_time.value
+            self.ptz_metrics[camera_name].stop_time.value = 0
+
        move_request = self.cams[camera_name]["relative_move_request"]

        # function takes in -1 to 1 for pan and tilt, interpolate to the values of the camera.
@@ -115,9 +115,11 @@ class PendingReviewSegment:
        if self._frame is not None:
            self.thumb_time = datetime.datetime.now().timestamp()
            self.has_frame = True
-            cv2.imwrite(
+            Path(self.frame_path).parent.mkdir(parents=True, exist_ok=True)
+            if not cv2.imwrite(
                self.frame_path, self._frame, [int(cv2.IMWRITE_WEBP_QUALITY), 60]
-            )
+            ):
+                logger.error("Failed to write review thumbnail to %s", self.frame_path)

    def save_full_frame(self, camera_config: CameraConfig, frame: np.ndarray) -> None:
        color_frame = cv2.cvtColor(frame, cv2.COLOR_YUV2BGR_I420)
@@ -128,9 +130,11 @@ class PendingReviewSegment:

        if self._frame is not None:
            self.has_frame = True
-            cv2.imwrite(
+            Path(self.frame_path).parent.mkdir(parents=True, exist_ok=True)
+            if not cv2.imwrite(
                self.frame_path, self._frame, [int(cv2.IMWRITE_WEBP_QUALITY), 60]
-            )
+            ):
+                logger.error("Failed to write review thumbnail to %s", self.frame_path)

    def get_data(self, ended: bool) -> dict:
        end_time = None
@@ -374,6 +378,16 @@ class ReviewSegmentMaintainer(threading.Thread):
        """Forcibly end the pending segment for a camera."""
        segment = self.active_review_segments.get(camera)
        if segment:
+            if self.indefinite_events.get(camera):
+                self.indefinite_events[camera] = {}
+                now = datetime.datetime.now().timestamp()
+
+                if segment.last_alert_time == sys.maxsize:
+                    segment.last_alert_time = now
+
+                if segment.last_detection_time == sys.maxsize:
+                    segment.last_detection_time = now
+
            prev_data = segment.get_data(False)
            return self._publish_segment_end(segment, prev_data)
        return None
@@ -0,0 +1,174 @@
+"""Tests that the internal port trusted by /auth cannot be moved at runtime."""
+
+import os
+import tempfile
+import unittest
+from unittest.mock import MagicMock, Mock, patch
+
+import ruamel.yaml
+from fastapi import Request
+
+from frigate.api.auth import get_allowed_cameras_for_filter, get_current_user
+from frigate.api.fastapi_app import create_fastapi_app
+from frigate.config import FrigateConfig
+from frigate.config.camera.updater import CameraConfigUpdatePublisher
+from frigate.const import JWT_SECRET_ENV_VAR
+from frigate.models import Event, Recordings, ReviewSegment
+from frigate.test.http_api.base_http_test import AuthTestClient, BaseTestHttp
+
+
+@patch.dict(os.environ, {JWT_SECRET_ENV_VAR: "test-secret"})
+class TestAuthInternalPort(BaseTestHttp):
+    """/auth grants anonymous admin by port, so that port must stay put.
+
+    nginx binds its listeners once at container start and never reloads them,
+    but /api/config/set can swap the live config object mid-process. If /auth
+    read the port off the live config, saving networking.listen.internal would
+    hand unauthenticated admin to whoever can reach the external port.
+    """
+
+    def setUp(self):
+        super().setUp(models=[Event, Recordings, ReviewSegment])
+        self.minimal_config = {
+            "mqtt": {"host": "mqtt"},
+            "auth": {"enabled": True},
+            "networking": {"listen": {"internal": 5000, "external": 8971}},
+            "cameras": {
+                "front_door": {
+                    "ffmpeg": {
+                        "inputs": [
+                            {"path": "rtsp://10.0.0.1:554/video", "roles": ["detect"]}
+                        ]
+                    },
+                    "detect": {
+                        "height": 1080,
+                        "width": 1920,
+                        "fps": 5,
+                    },
+                }
+            },
+        }
+
+    def _create_app(self):
+        mock_publisher = Mock(spec=CameraConfigUpdatePublisher)
+        mock_publisher.publisher = MagicMock()
+
+        app = create_fastapi_app(
+            FrigateConfig(**self.minimal_config),
+            self.db,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            mock_publisher,
+            None,
+            enforce_default_admin=False,
+        )
+
+        async def mock_get_current_user(request: Request):
+            return {
+                "username": request.headers.get("remote-user"),
+                "role": request.headers.get("remote-role"),
+            }
+
+        async def mock_get_allowed_cameras_for_filter(request: Request):
+            return list(self.minimal_config.get("cameras", {}).keys())
+
+        app.dependency_overrides[get_current_user] = mock_get_current_user
+        app.dependency_overrides[get_allowed_cameras_for_filter] = (
+            mock_get_allowed_cameras_for_filter
+        )
+
+        return app
+
+    def _write_config_file(self):
+        """Write the minimal config to a temp YAML file and return the path."""
+        yaml = ruamel.yaml.YAML()
+        f = tempfile.NamedTemporaryFile(mode="w", suffix=".yml", delete=False)
+        yaml.dump(self.minimal_config, f)
+        f.close()
+        return f.name
+
+    def test_internal_port_is_anonymous_admin(self):
+        app = self._create_app()
+
+        with AuthTestClient(app) as client:
+            resp = client.get("/auth", headers={"x-server-port": "5000"})
+
+            self.assertEqual(resp.status_code, 202)
+            self.assertEqual(resp.headers["remote-user"], "anonymous")
+            self.assertEqual(resp.headers["remote-role"], "admin")
+
+    def test_external_port_requires_auth(self):
+        app = self._create_app()
+
+        with AuthTestClient(app) as client:
+            resp = client.get("/auth", headers={"x-server-port": "8971"})
+
+            self.assertEqual(resp.status_code, 401)
+
+    def test_swapped_config_does_not_move_the_trusted_port(self):
+        """The live config is not what /auth trusts.
+
+        Stands in for every path that can rebind app.frigate_config while the
+        process runs, whatever restart flag the caller claimed.
+        """
+        app = self._create_app()
+
+        swapped = FrigateConfig(
+            **{
+                **self.minimal_config,
+                "networking": {"listen": {"internal": 8971, "external": 5000}},
+            }
+        )
+        app.frigate_config = swapped
+
+        with AuthTestClient(app) as client:
+            resp = client.get("/auth", headers={"x-server-port": "8971"})
+            self.assertEqual(resp.status_code, 401)
+
+            # nginx is still listening where it was told to at boot
+            resp = client.get("/auth", headers={"x-server-port": "5000"})
+            self.assertEqual(resp.status_code, 202)
+            self.assertEqual(resp.headers["remote-role"], "admin")
+
+    @patch("frigate.api.app.find_config_file")
+    def test_config_set_rejects_internal_matching_external(self, mock_find_config):
+        """Saving the internal port onto the external one is refused outright."""
+        config_path = self._write_config_file()
+        mock_find_config.return_value = config_path
+
+        try:
+            app = self._create_app()
+
+            with AuthTestClient(app) as client:
+                resp = client.put(
+                    "/config/set",
+                    json={
+                        "config_data": {"networking": {"listen": {"internal": 8971}}},
+                        "update_topic": "config/networking",
+                        "requires_restart": 1,
+                    },
+                )
+
+                self.assertEqual(resp.status_code, 400)
+                self.assertFalse(resp.json()["success"])
+
+                # the rejected save must not have reached the live config
+                self.assertEqual(
+                    app.frigate_config.networking.listen.internal_port, 5000
+                )
+
+                resp = client.get("/auth", headers={"x-server-port": "8971"})
+                self.assertEqual(resp.status_code, 401)
+
+            with open(config_path) as f:
+                self.assertNotIn("8971", f.read().split("external")[0])
+        finally:
+            os.unlink(config_path)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)
@@ -13,6 +13,7 @@ from frigate.config.camera.updater import (
    CameraConfigUpdatePublisher,
    CameraConfigUpdateTopic,
 )
+from frigate.config.holder import ConfigHolder
 from frigate.models import Event, Recordings, ReviewSegment
 from frigate.test.http_api.base_http_test import AuthTestClient, BaseTestHttp

@@ -373,6 +374,128 @@ class TestConfigSetWildcardPropagation(BaseTestHttp):
        finally:
            os.unlink(config_path)

+    @patch("frigate.api.app.find_config_file")
+    def test_global_birdseye_save_fans_out_resolved_camera_configs(
+        self, mock_find_config
+    ):
+        """A global birdseye save must also publish the per-camera values.
+
+        Global birdseye only seeds enabled and mode; the camera copies are what
+        the output process actually reads. Sending just the global object makes
+        a worker guess which cameras were inheriting, and the only available
+        guess (mode still equals the previous global) wrongly claims a camera
+        whose explicit yaml mode happens to match.
+        """
+        self.minimal_config["birdseye"] = {"enabled": True, "mode": "motion"}
+        # explicit override that matches the global value being replaced
+        self.minimal_config["cameras"]["front_door"]["birdseye"] = {"mode": "motion"}
+
+        config_path = self._write_config_file()
+        mock_find_config.return_value = config_path
+
+        try:
+            app, mock_publisher = self._create_app_with_publisher()
+            with AuthTestClient(app) as client:
+                resp = client.put(
+                    "/config/set",
+                    json={
+                        "config_data": {"birdseye": {"mode": "continuous"}},
+                        "update_topic": "config/birdseye",
+                        "requires_restart": 0,
+                    },
+                )
+
+                self.assertEqual(resp.status_code, 200)
+
+                # the global object still goes out on its own topic
+                mock_publisher.publisher.publish.assert_called_once()
+                topic, settings = mock_publisher.publisher.publish.call_args[0]
+                self.assertEqual(topic, "config/birdseye")
+                self.assertEqual(settings.mode.value, "continuous")
+
+                published = {
+                    call[0][0].camera: call[0][1]
+                    for call in mock_publisher.publish_update.call_args_list
+                }
+                self.assertEqual(set(published), {"front_door", "back_yard"})
+
+                for call in mock_publisher.publish_update.call_args_list:
+                    self.assertEqual(
+                        call[0][0].update_type, CameraConfigUpdateEnum.birdseye
+                    )
+
+                # the override survives, the inheriting camera follows global
+                self.assertEqual(published["front_door"].mode.value, "motion")
+                self.assertEqual(published["back_yard"].mode.value, "continuous")
+        finally:
+            os.unlink(config_path)
+
+    @patch("frigate.api.app.find_config_file")
+    def test_save_updates_the_config_holder(self, mock_find_config):
+        """A save must move the holder onto the freshly parsed config.
+
+        FrigateApp reads the holder when the watchdog rebuilds a crashed
+        process; if the save leaves it on the boot config, that process comes
+        back having lost every change made since Frigate started.
+        """
+        from fastapi import Request
+
+        from frigate.api.auth import get_allowed_cameras_for_filter, get_current_user
+        from frigate.api.fastapi_app import create_fastapi_app
+
+        config_path = self._write_config_file()
+        mock_find_config.return_value = config_path
+
+        mock_publisher = Mock(spec=CameraConfigUpdatePublisher)
+        mock_publisher.publisher = MagicMock()
+        boot_config = FrigateConfig(**self.minimal_config)
+        holder = ConfigHolder(boot_config)
+
+        try:
+            app = create_fastapi_app(
+                boot_config,
+                self.db,
+                None,
+                None,
+                None,
+                None,
+                None,
+                None,
+                mock_publisher,
+                None,
+                enforce_default_admin=False,
+                config_holder=holder,
+            )
+
+            async def mock_get_current_user(request: Request):
+                return {"username": "admin", "role": "admin"}
+
+            async def mock_get_allowed_cameras_for_filter(request: Request):
+                return list(self.minimal_config.get("cameras", {}).keys())
+
+            app.dependency_overrides[get_current_user] = mock_get_current_user
+            app.dependency_overrides[get_allowed_cameras_for_filter] = (
+                mock_get_allowed_cameras_for_filter
+            )
+
+            with AuthTestClient(app) as client:
+                resp = client.put(
+                    "/config/set",
+                    json={
+                        "config_data": {"birdseye": {"inactivity_threshold": 5}},
+                        "update_topic": "config/birdseye",
+                        "requires_restart": 0,
+                    },
+                )
+
+                self.assertEqual(resp.status_code, 200)
+
+            self.assertIsNot(holder.config, boot_config)
+            self.assertIs(holder.config, app.frigate_config)
+            self.assertEqual(holder.config.birdseye.inactivity_threshold, 5)
+        finally:
+            os.unlink(config_path)
+

 if __name__ == "__main__":
    unittest.main()
@@ -4,6 +4,7 @@ import unittest
 from unittest.mock import MagicMock

 from frigate.api.config_util import swap_runtime_config
+from frigate.config.holder import ConfigHolder


 class TestSwapRuntimeConfig(unittest.TestCase):
@@ -12,6 +13,7 @@ class TestSwapRuntimeConfig(unittest.TestCase):
    def _make_app(self) -> MagicMock:
        app = MagicMock()
        app.dispatcher.comms = [MagicMock(), MagicMock()]
+        app.config_holder = ConfigHolder(MagicMock(name="boot_config"))
        return app

    def test_rebinds_all_references(self) -> None:
@@ -37,11 +39,40 @@ class TestSwapRuntimeConfig(unittest.TestCase):
        # the swap rebuilds cameras from yaml, so overrides must be re-layered
        app.dispatcher.reapply_runtime_state_to_config.assert_called_once_with()

+    def test_updates_the_config_holder(self) -> None:
+        app = self._make_app()
+        holder = app.config_holder
+        config = MagicMock(name="new_config")
+
+        swap_runtime_config(app, config)
+
+        self.assertIs(holder.config, config)
+
+    def test_deferred_factory_builds_from_the_swapped_config(self) -> None:
+        """A watchdog-style factory must not rebuild from the boot config.
+
+        The factories in FrigateApp are lambdas evaluated when a process is
+        restarted, long after a user may have saved. Reading through the
+        holder is what keeps a rebuilt process from reverting every change
+        made since Frigate started.
+        """
+        app = self._make_app()
+        holder = app.config_holder
+        boot_config = holder.config
+        factory = lambda: holder.config  # noqa: E731
+        self.assertIs(factory(), boot_config)
+
+        config = MagicMock(name="new_config")
+        swap_runtime_config(app, config)
+
+        self.assertIs(factory(), config)
+
    def test_tolerates_missing_optional_collaborators(self) -> None:
        app = MagicMock()
        app.profile_manager = None
        app.stats_emitter = None
        app.dispatcher = None
+        app.config_holder = None
        config = MagicMock(name="new_config")

        # must not raise when the optional collaborators are absent
@@ -5,6 +5,7 @@ import tempfile
 import unittest
 from unittest.mock import MagicMock, patch

+from frigate.app import FrigateApp
 from frigate.comms.dispatcher import Dispatcher
 from frigate.comms.runtime_state import RuntimeStatePersistence

@@ -363,5 +364,94 @@ class TestReapplyRuntimeStateToConfig(unittest.TestCase):
        dispatcher.reapply_runtime_state_to_config()


+class TestStartupAppliesConfigLayersBeforeWorkersStart(unittest.TestCase):
+    """Both layers must reach the config before config-carrying workers start.
+
+    A worker started before a layer is applied keeps the yaml value for the
+    rest of the session: the config_updater broadcast sent later is dropped
+    for subscribers that have not connected yet, and nothing re-sends it.
+    """
+
+    CONFIG_LAYERS = (
+        "profile_manager.restore_persisted_profile_to_config",
+        "dispatcher.reapply_runtime_state_to_config",
+    )
+
+    # started with a copy of the camera config
+    CONFIG_CARRYING_WORKERS = (
+        "start_video_output_processor",
+        "start_ptz_autotracker",
+        "start_detected_frames_processor",
+        "start_camera_processor",
+        "start_audio_processor",
+    )
+
+    def _start_call_order(self) -> list[str]:
+        """Return the names FrigateApp.start() calls, in order."""
+        app = MagicMock()
+
+        with (
+            patch("frigate.app.set_file_limit"),
+            patch("frigate.app.cleanup_replay_cameras"),
+            patch("frigate.app.reap_stale_exports"),
+            patch("frigate.app.create_fastapi_app"),
+            patch("frigate.app.uvicorn"),
+        ):
+            FrigateApp.start(app)
+
+        return [name for name, _, _ in app.mock_calls]
+
+    def test_applied_before_any_config_carrying_worker(self) -> None:
+        order = self._start_call_order()
+
+        for layer in self.CONFIG_LAYERS:
+            for worker in self.CONFIG_CARRYING_WORKERS:
+                self.assertLess(order.index(layer), order.index(worker))
+
+    def test_applied_after_the_dispatcher_exists(self) -> None:
+        order = self._start_call_order()
+
+        for layer in self.CONFIG_LAYERS:
+            self.assertLess(order.index("init_dispatcher"), order.index(layer))
+
+    def test_applied_after_the_profile_base_is_snapshotted(self) -> None:
+        # ProfileManager snapshots the config as the "no profile" base that
+        # deactivation resets to, so neither layer may be in the config yet
+        order = self._start_call_order()
+
+        for layer in self.CONFIG_LAYERS:
+            self.assertLess(order.index("init_profile_manager"), order.index(layer))
+
+    def test_layers_applied_in_order(self) -> None:
+        # a runtime toggle is the layer the user set last, so it goes on top
+        order = self._start_call_order()
+
+        self.assertLess(
+            order.index("profile_manager.restore_persisted_profile_to_config"),
+            order.index("dispatcher.reapply_runtime_state_to_config"),
+        )
+
+    def test_overrides_still_re_applied_after_the_profile_is_restored(self) -> None:
+        # activation resets the sections it owns to the base first, so the
+        # overrides have to land on top again
+        order = self._start_call_order()
+
+        self.assertLess(
+            order.index("profile_manager.restore_persisted_profile"),
+            order.index("dispatcher.restore_runtime_state"),
+        )
+
+    def test_broadcast_replay_still_runs_at_the_end(self) -> None:
+        # the broadcast is the only channel for the recording, review, and
+        # embeddings processes, which start before the config can be corrected
+        order = self._start_call_order()
+
+        for replay in (
+            "profile_manager.restore_persisted_profile",
+            "dispatcher.restore_runtime_state",
+        ):
+            self.assertLess(order.index("start_audio_processor"), order.index(replay))
+
+
 if __name__ == "__main__":
    unittest.main()
@@ -491,6 +491,34 @@ class TestLlamaCppProvider(unittest.TestCase):
        final = _final_message(self._run_with_lines(client, lines, MULTIMODAL_MESSAGES))
        self.assertEqual(final["content"], "ok")

+    def _validated_client(self, server_context_size, provider_options=None):
+        """Build a client as if the server reported the given context size."""
+        cfg = GenAIConfig(
+            provider="llamacpp",
+            model="m",
+            base_url="http://localhost:9999",
+            provider_options=provider_options or {},
+        )
+        info = {
+            "context_size": server_context_size,
+            "supports_vision": False,
+            "supports_audio": False,
+            "supports_tools": False,
+            "supports_reasoning": False,
+            "media_marker": "<__media__>",
+        }
+        cls = PROVIDERS[GenAIProviderEnum.llamacpp]
+        with patch.object(cls, "_get_model_info", return_value=info):
+            return cls(cfg, timeout=5)
+
+    def test_server_context_size_used_without_override(self):
+        client = self._validated_client(4096)
+        self.assertEqual(client.get_context_size(), 4096)
+
+    def test_provider_options_context_size_overrides_server(self):
+        client = self._validated_client(4096, {"context_size": 32768})
+        self.assertEqual(client.get_context_size(), 32768)
+

 if __name__ == "__main__":
    unittest.main()
@@ -0,0 +1,41 @@
+"""Tests for networking config validation."""
+
+import unittest
+
+from pydantic import ValidationError
+
+from frigate.config.network import ListenConfig
+
+
+class TestListenConfig(unittest.TestCase):
+    def test_defaults_are_distinct(self):
+        listen = ListenConfig()
+
+        self.assertEqual(listen.internal_port, 5000)
+        self.assertEqual(listen.external_port, 8971)
+
+    def test_address_and_port_string_is_parsed(self):
+        listen = ListenConfig(internal="127.0.0.1:5000", external="0.0.0.0:8971")
+
+        self.assertEqual(listen.internal_port, 5000)
+        self.assertEqual(listen.external_port, 8971)
+
+    def test_identical_ports_rejected(self):
+        with self.assertRaises(ValidationError):
+            ListenConfig(internal=8971, external=8971)
+
+    def test_same_port_on_different_addresses_rejected(self):
+        # nginx would accept these as distinct listeners, but /auth decides on
+        # the port alone, so the external one would inherit anonymous admin
+        with self.assertRaises(ValidationError):
+            ListenConfig(internal="127.0.0.1:8971", external="0.0.0.0:8971")
+
+    def test_distinct_ports_accepted(self):
+        listen = ListenConfig(internal=5001, external="0.0.0.0:8971")
+
+        self.assertEqual(listen.internal_port, 5001)
+        self.assertEqual(listen.external_port, 8971)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)
@@ -0,0 +1,226 @@
+"""Tests for detector post-processing NMS box format handling.
+
+cv2.dnn.NMSBoxes expects boxes as [x, y, width, height]. Passing corner
+coordinates [x1, y1, x2, y2] makes OpenCV treat x2/y2 as width/height,
+inflating every box toward the bottom-right by its distance from the origin.
+Two well separated objects far from the origin then appear to overlap and the
+lower scoring one is silently suppressed.
+
+The regression geometry used throughout: two boxes with zero true overlap,
+A = (393, 499, 484, 620) and B = (527, 499, 618, 620) in a 640x640 input
+(43 px gap). Misread as [x, y, w, h] their IoU is 0.465, above the 0.4 NMS
+threshold, so the buggy format drops the lower scoring box while correct
+conversion keeps both.
+"""
+
+import math
+import unittest
+from queue import Queue
+
+import numpy as np
+
+from frigate.detectors.plugins.memryx import MemryXDetector
+from frigate.util.model import (
+    post_process_dfine,
+    post_process_rfdetr,
+    post_process_yolo,
+    post_process_yolox,
+)
+
+WIDTH = 640
+HEIGHT = 640
+
+# box A: xyxy (393, 499, 484, 620) as center format
+A_CX, A_CY, A_W, A_H = 438.5, 559.5, 91.0, 121.0
+# box B: xyxy (527, 499, 618, 620) as center format
+B_CX, B_CY, B_W, B_H = 572.5, 559.5, 91.0, 121.0
+
+# expected normalized output rows: [class_id, conf, y1, x1, y2, x2]
+A_ROW = [499 / 640, 393 / 640, 620 / 640, 484 / 640]
+B_ROW = [499 / 640, 527 / 640, 620 / 640, 618 / 640]
+
+
+def kept(detections: np.ndarray) -> np.ndarray:
+    """Rows of the padded (20, 6) output that hold real detections."""
+    return detections[detections[:, 1] > 0]
+
+
+class TestYoloNmsPostProcess(unittest.TestCase):
+    def _single_output(self, rows: list[list[float]]) -> list[np.ndarray]:
+        """Build a single-tensor YOLO output (1, attrs, anchors) from
+        [cx, cy, w, h, class scores...] rows, padded with empty anchors."""
+        anchors = np.zeros((10, len(rows[0])), dtype=np.float32)
+        anchors[: len(rows)] = np.array(rows, dtype=np.float32)
+        return [anchors.T[np.newaxis, ...]]
+
+    def test_keeps_separated_objects_far_from_origin(self):
+        output = self._single_output(
+            [
+                [A_CX, A_CY, A_W, A_H, 0.90, 0.0],
+                [B_CX, B_CY, B_W, B_H, 0.0, 0.85],
+            ]
+        )
+
+        detections = kept(post_process_yolo(output, WIDTH, HEIGHT))
+
+        self.assertEqual(len(detections), 2)
+        np.testing.assert_allclose(detections[0], [0, 0.90, *A_ROW], atol=2e-3)
+        np.testing.assert_allclose(detections[1], [1, 0.85, *B_ROW], atol=2e-3)
+
+    def test_still_suppresses_true_duplicates(self):
+        # same object twice, shifted 4 px: true IoU 0.92, must dedupe to one
+        output = self._single_output(
+            [
+                [A_CX, A_CY, A_W, A_H, 0.90, 0.0],
+                [A_CX + 4, A_CY, A_W, A_H, 0.85, 0.0],
+            ]
+        )
+
+        detections = kept(post_process_yolo(output, WIDTH, HEIGHT))
+
+        self.assertEqual(len(detections), 1)
+        np.testing.assert_allclose(detections[0], [0, 0.90, *A_ROW], atol=2e-3)
+
+
+class TestMultipartYoloPostProcess(unittest.TestCase):
+    def _multipart_output(self) -> list[np.ndarray]:
+        """Build a 3-scale anchor-based YOLO output containing boxes A and B,
+        both decoded through anchor 0 of the stride-32 scale."""
+        outputs = [
+            np.zeros((1, 255, 80, 80), dtype=np.float32),
+            np.zeros((1, 255, 40, 40), dtype=np.float32),
+            np.zeros((1, 255, 20, 20), dtype=np.float32),
+        ]
+        stride, (anchor_w, anchor_h) = 32, (142, 110)
+
+        for cx, cy, w, h, conf, class_channel in [
+            (A_CX, A_CY, A_W, A_H, 0.95, 5),  # class 0
+            (B_CX, B_CY, B_W, B_H, 0.90, 6),  # class 1
+        ]:
+            cell_x, cell_y = int(cx // stride), int(cy // stride)
+            dx = (cx / stride - cell_x + 0.5) / 2
+            dy = (cy / stride - cell_y + 0.5) / 2
+            dw = math.sqrt(w / anchor_w) / 2
+            dh = math.sqrt(h / anchor_h) / 2
+            # anchor 0 occupies channels 0-84 of the 255 channel tensor
+            outputs[2][0, 0:4, cell_y, cell_x] = [dx, dy, dw, dh]
+            outputs[2][0, 4, cell_y, cell_x] = conf
+            outputs[2][0, class_channel, cell_y, cell_x] = 1.0
+
+        return outputs
+
+    def test_keeps_separated_objects_far_from_origin(self):
+        detections = kept(post_process_yolo(self._multipart_output(), WIDTH, HEIGHT))
+
+        self.assertEqual(len(detections), 2)
+        np.testing.assert_allclose(detections[0], [0, 0.95, *A_ROW], atol=2e-3)
+        np.testing.assert_allclose(detections[1], [1, 0.90, *B_ROW], atol=2e-3)
+
+    def test_empty_output_returns_no_detections(self):
+        outputs = [
+            np.zeros((1, 255, 80, 80), dtype=np.float32),
+            np.zeros((1, 255, 40, 40), dtype=np.float32),
+            np.zeros((1, 255, 20, 20), dtype=np.float32),
+        ]
+
+        detections = kept(post_process_yolo(outputs, WIDTH, HEIGHT))
+
+        self.assertEqual(len(detections), 0)
+
+
+class TestYoloxPostProcess(unittest.TestCase):
+    def test_keeps_separated_objects_far_from_origin(self):
+        # with zero grids and unit strides the decode reduces to
+        # cx = raw cx and w = exp(raw w)
+        rows = np.zeros((10, 7), dtype=np.float32)
+        rows[0] = [A_CX, A_CY, math.log(A_W), math.log(A_H), 1.0, 0.90, 0.0]
+        rows[1] = [B_CX, B_CY, math.log(B_W), math.log(B_H), 1.0, 0.0, 0.85]
+        predictions = rows[np.newaxis, ...]
+        grids = np.zeros((1, 10, 2), dtype=np.float32)
+        expanded_strides = np.ones((1, 10, 1), dtype=np.float32)
+
+        detections = kept(
+            post_process_yolox(predictions, WIDTH, HEIGHT, grids, expanded_strides)
+        )
+
+        self.assertEqual(len(detections), 2)
+        np.testing.assert_allclose(detections[0], [0, 0.90, *A_ROW], atol=2e-3)
+        np.testing.assert_allclose(detections[1], [1, 0.85, *B_ROW], atol=2e-3)
+
+
+class TestDfinePostProcess(unittest.TestCase):
+    def test_keeps_separated_objects_far_from_origin(self):
+        # D-FINE emits absolute pixel xyxy boxes alongside labels and scores
+        labels = np.zeros((1, 10), dtype=np.int64)
+        labels[0, 1] = 1
+        boxes = np.zeros((1, 10, 4), dtype=np.float32)
+        boxes[0, 0] = [393, 499, 484, 620]
+        boxes[0, 1] = [527, 499, 618, 620]
+        scores = np.zeros((1, 10), dtype=np.float32)
+        scores[0, 0] = 0.90
+        scores[0, 1] = 0.85
+
+        detections = kept(post_process_dfine([labels, boxes, scores], WIDTH, HEIGHT))
+
+        self.assertEqual(len(detections), 2)
+        np.testing.assert_allclose(detections[0], [0, 0.90, *A_ROW], atol=2e-3)
+        np.testing.assert_allclose(detections[1], [1, 0.85, *B_ROW], atol=2e-3)
+
+
+class TestRfdetrPostProcess(unittest.TestCase):
+    def test_keeps_separated_objects_far_from_origin(self):
+        # RF-DETR emits normalized center format boxes and class logits where
+        # logit index 0 is the background class
+        boxes = np.zeros((1, 10, 4), dtype=np.float32)
+        boxes[0, 0] = [A_CX / WIDTH, A_CY / HEIGHT, A_W / WIDTH, A_H / HEIGHT]
+        boxes[0, 1] = [B_CX / WIDTH, B_CY / HEIGHT, B_W / WIDTH, B_H / HEIGHT]
+        # background heavy logits everywhere, then two confident objects
+        logits = np.tile(np.array([10.0, 0.0, 0.0], dtype=np.float32), (1, 10, 1))
+        logits[0, 0] = [0.0, 4.0, 0.0]  # class 0 after background offset
+        logits[0, 1] = [0.0, 0.0, 3.5]  # class 1 after background offset
+
+        detections = kept(post_process_rfdetr([boxes, logits]))
+
+        conf_a = math.exp(4.0) / (math.exp(4.0) + 2)
+        conf_b = math.exp(3.5) / (math.exp(3.5) + 2)
+        self.assertEqual(len(detections), 2)
+        np.testing.assert_allclose(detections[0], [0, conf_a, *A_ROW], atol=2e-3)
+        np.testing.assert_allclose(detections[1], [1, conf_b, *B_ROW], atol=2e-3)
+
+
+class TestMemryxSsdlitePostProcess(unittest.TestCase):
+    def test_keeps_separated_objects_far_from_origin(self):
+        # the NMS math runs on the host CPU, so the real method is testable
+        # without MemryX hardware; it only needs the model dimensions and
+        # the output queue
+        detector = object.__new__(MemryXDetector)
+        detector.memx_model_width = WIDTH
+        detector.memx_model_height = HEIGHT
+        detector.output_queue = Queue()
+
+        # this path uses a 0.5 NMS threshold, so use a tighter pair: zero
+        # true overlap (10 px gap), IoU 0.69 when misread as [x, y, w, h]
+        dets = np.zeros((1, 10, 5), dtype=np.float32)
+        dets[0, 0] = [480, 500, 540, 620, 0.90]
+        dets[0, 1] = [550, 500, 610, 620, 0.85]
+        labels = np.zeros((1, 10), dtype=np.float32)
+        labels[0, 1] = 1
+
+        detector.post_process_ssdlite([dets, labels])
+        detections = kept(detector.output_queue.get())
+
+        self.assertEqual(len(detections), 2)
+        np.testing.assert_allclose(
+            detections[0],
+            [0, 0.90, 500 / 640, 480 / 640, 620 / 640, 540 / 640],
+            atol=2e-3,
+        )
+        np.testing.assert_allclose(
+            detections[1],
+            [1, 0.85, 500 / 640, 550 / 640, 620 / 640, 610 / 640],
+            atol=2e-3,
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -785,6 +785,98 @@ class TestProfileManager(unittest.TestCase):
        manager.activate_profile("armed", clear_runtime_overrides=False)
        dispatcher.clear_runtime_state.assert_not_called()

+    def test_apply_profile_to_config_mutates_the_config(self):
+        """The config-only half applies the same overrides as activation."""
+        err = self.manager.apply_profile_to_config("armed")
+        assert err is None
+
+        front = self.config.cameras["front"]
+        assert front.notifications.enabled is True
+        assert front.objects.track == ["person", "car", "package"]
+
+    def test_apply_profile_to_config_makes_no_zmq_mqtt_or_disk_writes(self):
+        """Workers are started with the values, so nothing is published yet."""
+        dispatcher = MagicMock()
+        manager = ProfileManager(self.config, self.mock_updater, dispatcher)
+
+        with patch.object(ProfileManager, "_persist_active_profile") as mock_persist:
+            manager.apply_profile_to_config("armed")
+
+        self.mock_updater.publish_update.assert_not_called()
+        dispatcher.publish.assert_not_called()
+        mock_persist.assert_not_called()
+        # bookkeeping stays with activate_profile
+        assert self.config.active_profile is None
+
+    def test_apply_profile_to_config_rejects_an_unknown_profile(self):
+        err = self.manager.apply_profile_to_config("nonexistent")
+        assert err is not None
+        assert "not defined" in err
+
+    def test_restore_persisted_profile_to_config_applies_it(self):
+        """The startup config pass restores what was persisted."""
+        with patch.object(
+            ProfileManager, "load_persisted_profile", return_value="armed"
+        ):
+            self.manager.restore_persisted_profile_to_config()
+
+        assert self.config.cameras["front"].notifications.enabled is True
+        # still the config-only half, so nothing is published or persisted
+        self.mock_updater.publish_update.assert_not_called()
+        assert self.config.active_profile is None
+
+    def test_restore_persisted_profile_to_config_no_op_when_none_persisted(self):
+        with patch.object(ProfileManager, "load_persisted_profile", return_value=None):
+            self.manager.restore_persisted_profile_to_config()
+
+        assert self.config.cameras["front"].notifications.enabled is False
+
+    def test_restore_persisted_profile_to_config_ignores_a_stale_name(self):
+        """A profile no longer offered by any camera must not be applied."""
+        with patch.object(
+            ProfileManager, "load_persisted_profile", return_value="ghost"
+        ):
+            self.manager.restore_persisted_profile_to_config()
+
+        assert self.config.cameras["front"].notifications.enabled is False
+
+    @patch.object(ProfileManager, "_persist_active_profile")
+    def test_restore_persisted_profile_activates_and_publishes(self, mock_persist):
+        """The startup publish pass runs a full activation."""
+        dispatcher = MagicMock()
+        manager = ProfileManager(self.config, self.mock_updater, dispatcher)
+
+        with patch.object(
+            ProfileManager, "load_persisted_profile", return_value="armed"
+        ):
+            manager.restore_persisted_profile()
+
+        assert self.config.active_profile == "armed"
+        self.mock_updater.publish_update.assert_called()
+        # a startup replay must not wipe the runtime overrides layered on top
+        dispatcher.clear_runtime_state.assert_not_called()
+
+    @patch.object(ProfileManager, "_persist_active_profile")
+    def test_activation_after_apply_still_publishes_every_section(self, mock_persist):
+        """Re-deriving the same state must not skip the broadcast.
+
+        The processes that started before the config was corrected have no
+        other channel.
+        """
+        self.manager.apply_profile_to_config("armed")
+        self.mock_updater.publish_update.reset_mock()
+
+        err = self.manager.activate_profile("armed", clear_runtime_overrides=False)
+        assert err is None
+
+        published = {
+            call.args[0].update_type.name
+            for call in self.mock_updater.publish_update.call_args_list
+        }
+        assert "notifications" in published
+        assert "objects" in published
+        assert self.config.active_profile == "armed"
+
    @patch.object(ProfileManager, "_persist_active_profile")
    def test_update_config_preserves_runtime_state_with_active_profile(
        self, mock_persist
@@ -1,4 +1,4 @@
-"""Tests for ONVIF init state that must not depend on the autotracking config.
+"""Tests for ONVIF state that must not depend on the autotracking config.

 Regression coverage for a camera that is initialized while autotracking is off and
 has it enabled later, which is the normal wizard flow: set the camera up first,
@@ -10,12 +10,17 @@ the tracking thread.

 The request objects are built from the locally parsed WSDL and cost no network, so
 they are always created and init=True now implies they exist.
+
+Also covers the inverse direction: the ptz movement timestamps must not be written
+for a camera that has autotracking off, because nothing clears them back out.
 """

 import unittest
 from unittest.mock import AsyncMock, MagicMock

+from frigate.camera import PTZMetrics
 from frigate.config import FrigateConfig
+from frigate.ptz.autotrack import ptz_moving_at_frame_time
 from frigate.ptz.onvif import OnvifController

 CAMERA = "ptz_cam"
@@ -97,6 +102,36 @@ def _make_controller(autotracking_enabled: bool) -> OnvifController:
    return controller


+def _make_move_controller(autotracking_enabled: bool) -> OnvifController:
+    """Build an already initialized controller for a camera that supports relative
+    FOV movement, with real metrics so the timestamp writes can be asserted on."""
+    config = _config(autotracking_enabled)
+    controller = OnvifController.__new__(OnvifController)
+    controller.config = config
+    controller.camera_configs = {CAMERA: config.cameras[CAMERA]}
+    controller.failed_cams = {}
+
+    ptz = MagicMock()
+    ptz.RelativeMove = AsyncMock()
+    controller.cams = {
+        CAMERA: {
+            "init": True,
+            "active": False,
+            "ptz": ptz,
+            "features": ["pt", "pt-r-fov"],
+            "relative_move_request": MagicMock(),
+            "relative_fov_range": {
+                "XRange": {"Min": -1.0, "Max": 1.0},
+                "YRange": {"Min": -1.0, "Max": 1.0},
+            },
+        }
+    }
+    controller.ptz_metrics = {
+        CAMERA: PTZMetrics(autotracker_enabled=autotracking_enabled)
+    }
+    return controller
+
+
 class TestOnvifInitRequests(unittest.IsolatedAsyncioTestCase):
    async def test_status_request_created_when_autotracking_disabled(self) -> None:
        # the wizard flow: onvif configured first, autotracking enabled later
@@ -143,5 +178,61 @@ class TestOnvifInitRequests(unittest.IsolatedAsyncioTestCase):
        ptz.GetStatus.assert_not_called()


+class TestManualRelativeMoveMetrics(unittest.IsolatedAsyncioTestCase):
+    """A manual move from the UI (click to move, drag to zoom) sends move_relative
+    for any camera that advertises pt-r-fov, autotracking or not."""
+
+    async def test_metrics_untouched_when_autotracking_disabled(self) -> None:
+        # only camera_maintenance polls get_camera_status, and only for autotracking
+        # cameras, so a manual move that starts the clock here is never stopped
+        controller = _make_move_controller(autotracking_enabled=False)
+        metrics = controller.ptz_metrics[CAMERA]
+        metrics.frame_time.value = 1000.0
+
+        await controller._move_relative(CAMERA, 0.25, -0.25, 0, 1)
+
+        controller.cams[CAMERA]["ptz"].RelativeMove.assert_awaited_once()
+        self.assertEqual(metrics.start_time.value, 0)
+        self.assertEqual(metrics.stop_time.value, 0)
+        self.assertTrue(metrics.motor_stopped.is_set())
+
+    async def test_detection_regions_not_suppressed_after_manual_move(self) -> None:
+        # the symptom of the bug: object detection stops entirely because motion
+        # boxes are never promoted to detection regions again
+        controller = _make_move_controller(autotracking_enabled=False)
+        metrics = controller.ptz_metrics[CAMERA]
+        metrics.frame_time.value = 1000.0
+
+        await controller._move_relative(CAMERA, 0.25, -0.25, 0, 1)
+
+        for later_frame_time in (1001.0, 1060.0, 4600.0):
+            with self.subTest(frame_time=later_frame_time):
+                self.assertFalse(
+                    ptz_moving_at_frame_time(
+                        later_frame_time,
+                        metrics.start_time.value,
+                        metrics.stop_time.value,
+                    )
+                )
+
+    async def test_metrics_written_when_autotracking_enabled(self) -> None:
+        # get_camera_status resets stop_time once the camera reports IDLE, so the
+        # autotracking path keeps its motion estimation timestamps
+        controller = _make_move_controller(autotracking_enabled=True)
+        metrics = controller.ptz_metrics[CAMERA]
+        metrics.frame_time.value = 1000.0
+
+        await controller._move_relative(CAMERA, 0.25, -0.25, 0, 1)
+
+        self.assertEqual(metrics.start_time.value, 1000.0)
+        self.assertEqual(metrics.stop_time.value, 0)
+        self.assertFalse(metrics.motor_stopped.is_set())
+        self.assertTrue(
+            ptz_moving_at_frame_time(
+                1001.0, metrics.start_time.value, metrics.stop_time.value
+            )
+        )
+
+
 if __name__ == "__main__":
    unittest.main()
@@ -0,0 +1,63 @@
+"""Tests for embedding cleanup on the main Frigate database.
+
+Embeddings are deleted whether or not semantic search is currently enabled, so
+the delete path has to tolerate databases where the vec0 tables were never
+created and installs where the sqlite-vec extension is unavailable.
+"""
+
+import os
+import tempfile
+import unittest
+
+from frigate.db.sqlitevecq import SqliteVecQueueDatabase
+
+
+class TestDeleteEmbeddings(unittest.TestCase):
+    def setUp(self) -> None:
+        self.tmp_dir = tempfile.TemporaryDirectory()
+        self.db = SqliteVecQueueDatabase(os.path.join(self.tmp_dir.name, "test.db"))
+        self.db.start()
+        # the extension is not available to tests, so stand in for a database
+        # that has it loaded and use a plain table for the deletes
+        self.db.load_vec_extension = True
+
+    def tearDown(self) -> None:
+        self.db.stop()
+        self.db.close()
+        self.tmp_dir.cleanup()
+
+    def _flush_writes(self) -> None:
+        # writes are queued and applied by a worker thread, and the queue is
+        # FIFO, so awaiting a later write means the earlier ones are done
+        self.db.execute_sql("PRAGMA user_version = 0").fetchall()
+
+    def _create_thumbnails_table(self) -> None:
+        self.db.execute_sql("CREATE TABLE vec_thumbnails (id TEXT PRIMARY KEY)")
+        self.db.execute_sql("INSERT INTO vec_thumbnails (id) VALUES ('a'), ('b')")
+        self._flush_writes()
+
+    def _thumbnail_ids(self) -> list[str]:
+        return [row[0] for row in self.db.execute_sql("SELECT id FROM vec_thumbnails")]
+
+    def test_delete_without_tables_does_not_raise(self) -> None:
+        # semantic search was never enabled, so event cleanup has nothing to do
+        self.db.delete_embeddings_thumbnail(event_ids=["1700000000.0-abc"])
+        self.db.delete_embeddings_description(event_ids=["1700000000.0-abc"])
+
+    def test_delete_removes_embeddings(self) -> None:
+        self._create_thumbnails_table()
+
+        self.db.delete_embeddings_thumbnail(event_ids=["a"])
+        self._flush_writes()
+
+        self.assertEqual(self._thumbnail_ids(), ["b"])
+
+    def test_delete_skipped_without_extension(self) -> None:
+        self._create_thumbnails_table()
+        self.db.load_vec_extension = False
+
+        self.db.delete_embeddings_thumbnail(event_ids=["a"])
+        self._flush_writes()
+
+        # the vec0 tables cannot be written without the extension
+        self.assertEqual(self._thumbnail_ids(), ["a", "b"])
@@ -472,6 +472,18 @@ def sanitize_float(value):
    return value


+def has_non_finite_number(value: Any) -> bool:
+    """Return True if any number in a parsed JSON value is NaN or infinite."""
+    if isinstance(value, float):
+        return not math.isfinite(value)
+    if isinstance(value, dict):
+        return any(has_non_finite_number(v) for v in value.values())
+    if isinstance(value, list):
+        return any(has_non_finite_number(v) for v in value)
+
+    return False
+
+
 def cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
    return 1 - cosine_distance(a, b)

@@ -16,6 +16,31 @@ logger = logging.getLogger(__name__)
 ### Post Processing


+def xyxy_to_xywh_for_nms(boxes: np.ndarray | list) -> np.ndarray:
+    """Convert [x1, y1, x2, y2] boxes to the [x, y, width, height] format
+    that cv2.dnn.NMSBoxes expects.
+
+    Passing corner coordinates directly makes OpenCV treat x2/y2 as the box
+    size, inflating every box toward the bottom-right by its distance from
+    the origin, which suppresses valid detections near other objects.
+
+    Args:
+        boxes: Array-like of shape (N, 4) in corner format.
+
+    Returns:
+        Float32 array of shape (N, 4) in top-left plus size format.
+    """
+    boxes = np.asarray(boxes, dtype=np.float32)
+
+    if boxes.size == 0:
+        return np.zeros((0, 4), dtype=np.float32)
+
+    xywh = boxes.copy()
+    xywh[:, 2] -= xywh[:, 0]
+    xywh[:, 3] -= xywh[:, 1]
+    return xywh
+
+
 def post_process_dfine(
    tensor_output: np.ndarray, width: int, height: int
 ) -> np.ndarray:
@@ -25,7 +50,9 @@ def post_process_dfine(

    input_shape = np.array([height, width, height, width])
    boxes = np.divide(boxes, input_shape, dtype=np.float32)
-    indices = cv2.dnn.NMSBoxes(boxes, scores, score_threshold=0.4, nms_threshold=0.4)
+    indices = cv2.dnn.NMSBoxes(
+        xyxy_to_xywh_for_nms(boxes), scores, score_threshold=0.4, nms_threshold=0.4
+    )
    detections = np.zeros((20, 6), np.float32)

    for i, (bbox, confidence, class_id) in enumerate(
@@ -78,7 +105,10 @@ def post_process_rfdetr(tensor_output: list[np.ndarray, np.ndarray]) -> np.ndarr

    # apply nms
    indices = cv2.dnn.NMSBoxes(
-        filtered_boxes, filtered_scores, score_threshold=0.4, nms_threshold=0.4
+        xyxy_to_xywh_for_nms(filtered_boxes),
+        filtered_scores,
+        score_threshold=0.4,
+        nms_threshold=0.4,
    )
    detections = np.zeros((20, 6), np.float32)

@@ -159,7 +189,7 @@ def __post_process_multipart_yolo(
                    all_class_ids.append(class_id)

    indices = cv2.dnn.NMSBoxes(
-        bboxes=all_boxes,
+        bboxes=xyxy_to_xywh_for_nms(all_boxes),
        scores=all_scores,
        score_threshold=0.4,
        nms_threshold=0.4,
@@ -206,7 +236,9 @@ def __post_process_nms_yolo(predictions: np.ndarray, width, height) -> np.ndarra
    boxes = boxes_xyxy

    # run NMS
-    indices = cv2.dnn.NMSBoxes(boxes, scores, score_threshold=0.4, nms_threshold=0.4)
+    indices = cv2.dnn.NMSBoxes(
+        xyxy_to_xywh_for_nms(boxes), scores, score_threshold=0.4, nms_threshold=0.4
+    )
    detections = np.zeros((20, 6), np.float32)
    for i, (bbox, confidence, class_id) in enumerate(
        zip(boxes[indices], scores[indices], class_ids[indices])
@@ -258,7 +290,7 @@ def post_process_yolox(
    scores = scores[np.arange(len(cls_inds)), cls_inds]

    indices = cv2.dnn.NMSBoxes(
-        boxes_xyxy, scores, score_threshold=0.4, nms_threshold=0.4
+        xyxy_to_xywh_for_nms(boxes_xyxy), scores, score_threshold=0.4, nms_threshold=0.4
    )

    detections = np.zeros((20, 6), np.float32)
@@ -326,7 +358,7 @@ def get_ort_providers(
                    {
                        "device_id": device_id,
                        "trt_fp16_enable": requires_fp16
-                        and os.environ.get("USE_FP_16", "True") != "False",
+                        and os.environ.get("USE_FP16", "True") != "False",
                        "trt_timing_cache_enable": True,
                        "trt_engine_cache_enable": True,
                        "trt_timing_cache_path": os.path.join(
@@ -35,6 +35,11 @@ logger = logging.getLogger(__name__)
 GRID_SIZE = 8


+def create_empty_regions_grid() -> list[list[dict[str, Any]]]:
+    """Create a region grid with no learned sizes."""
+    return [[{"sizes": []} for _ in range(GRID_SIZE)] for _ in range(GRID_SIZE)]
+
+
 def get_camera_regions_grid(
    name: str,
    detect: DetectConfig,
@@ -47,12 +52,7 @@ def get_camera_regions_grid(
        grid = regions.grid
        last_update = regions.last_update
    except DoesNotExist:
-        grid = []
-        for x in range(GRID_SIZE):
-            row = []
-            for y in range(GRID_SIZE):
-                row.append({"sizes": []})
-            grid.append(row)
+        grid = create_empty_regions_grid()
        last_update = 0

    # get events for timeline entries
@@ -358,12 +358,17 @@ def process_frames(
            ]

            # only add in the motion boxes when not calibrating and a ptz is not moving via autotracking
-            # ptz_moving_at_frame_time() always returns False for non-autotracking cameras
-            if not motion_detector.is_calibrating() and not ptz_moving_at_frame_time(
-                frame_time,
-                ptz_metrics.start_time.value,
-                ptz_metrics.stop_time.value,
-            ):
+            # the ptz timestamps are only maintained while autotracking is on, so gate
+            # on the metric rather than trusting them to be reset otherwise
+            ptz_moving = ptz_metrics.autotracker_enabled.value and (
+                ptz_moving_at_frame_time(
+                    frame_time,
+                    ptz_metrics.start_time.value,
+                    ptz_metrics.stop_time.value,
+                )
+            )
+
+            if not motion_detector.is_calibrating() and not ptz_moving:
                # find motion boxes that are not inside tracked object regions
                standalone_motion_boxes = [
                    b for b in motion_boxes if not inside_any(b, regions)
@@ -36,8 +36,8 @@ from __future__ import annotations
 import argparse
 import os
 import sys
+from collections.abc import Iterable
 from dataclasses import dataclass
-from typing import Iterable

 import cv2
 import numpy as np
@@ -53,21 +53,31 @@ ARCFACE_INPUT_SIZE = 112
 # ---------------------------------------------------------------------------


+def _bgr_to_rgb(frame: np.ndarray) -> np.ndarray:
+    """Mirror BaseEmbedding._bgr_to_rgb."""
+    if isinstance(frame, np.ndarray) and frame.ndim == 3:
+        return np.ascontiguousarray(frame[:, :, ::-1])
+
+    return frame
+
+
 def _process_image_frigate(image: np.ndarray) -> Image.Image:
    """Mirror BaseEmbedding._process_image for an ndarray input.

-    NOTE: Frigate passes the output of `cv2.imread` (BGR) directly in. PIL's
-    `Image.fromarray` does NOT reorder channels, so the embedder effectively
-    receives a BGR-ordered tensor. We replicate that faithfully here. (Tested
-    — swapping to RGB produces near-identical embeddings; this model is
-    robust to channel order.)
+    `Image.fromarray` does not reorder channels, so whatever order it is
+    handed is what reaches the model. Callers swap to RGB first, exactly as
+    ArcfaceEmbedding._preprocess_inputs does.
    """
    return Image.fromarray(image)


 def arcface_preprocess(image_bgr: np.ndarray) -> np.ndarray:
-    """Mirror ArcfaceEmbedding._preprocess_inputs."""
-    pil = _process_image_frigate(image_bgr)
+    """Mirror ArcfaceEmbedding._preprocess_inputs.
+
+    Face crops arrive BGR from cv2 and #23712 added the swap to RGB before
+    embedding, so this script has to do it too.
+    """
+    pil = _process_image_frigate(_bgr_to_rgb(image_bgr))

    width, height = pil.size
    if width != ARCFACE_INPUT_SIZE or height != ARCFACE_INPUT_SIZE:
@@ -138,9 +148,7 @@ class LandmarkAligner:
        M[0, 2] += tX - eyesCenter[0]
        M[1, 2] += tY - eyesCenter[1]

-        aligned = cv2.warpAffine(
-            image, M, (out_w, out_h), flags=cv2.INTER_CUBIC
-        )
+        aligned = cv2.warpAffine(image, M, (out_w, out_h), flags=cv2.INTER_CUBIC)
        info = dict(
            angle=float(angle),
            eye_dist_px=dist,
@@ -433,9 +441,7 @@ def vector_outlier_test(
        if neg
        else np.array([])
    )
-    baseline_conf_neg = np.array(
-        [similarity_to_confidence(c) for c in baseline_neg]
-    )
+    baseline_conf_neg = np.array([similarity_to_confidence(c) for c in baseline_neg])

    print(
        f"\nBaseline (trim_mean only, {len(pos)} images):"
@@ -465,9 +471,7 @@ def vector_outlier_test(
        mean, keep = iterative_mean(all_embs, T)
        pos_sims = np.array([cosine(p.embedding, mean) for p in pos])
        neg_sims = (
-            np.array([cosine(n.embedding, mean) for n in neg])
-            if neg
-            else np.array([])
+            np.array([cosine(n.embedding, mean) for n in neg]) if neg else np.array([])
        )
        neg_conf = np.array([similarity_to_confidence(c) for c in neg_sims])
        margin = pos_sims.min() - (neg_sims.max() if len(neg_sims) else 0)
@@ -483,9 +487,7 @@ def vector_outlier_test(
    # Show which images get dropped at the shipped threshold + neighbors
    for T_show in (0.25, 0.30, 0.33):
        _, keep = iterative_mean(all_embs, T_show)
-        print(
-            f"\nAt T={T_show}, the {int((~keep).sum())} dropped positives are:"
-        )
+        print(f"\nAt T={T_show}, the {int((~keep).sum())} dropped positives are:")
        final_mean = stats.trim_mean(all_embs[keep], base_trim, axis=0)
        m_n = final_mean / (np.linalg.norm(final_mean) + 1e-9)
        for i, (p, k) in enumerate(zip(pos, keep)):
@@ -501,9 +503,7 @@ def vector_outlier_test(
                )


-def degenerate_embedding_test(
-    pos: list[FaceSample], neg: list[FaceSample]
-) -> None:
+def degenerate_embedding_test(pos: list[FaceSample], neg: list[FaceSample]) -> None:
    """Detect whether negatives and low-quality positives share a degenerate
    'tiny/noisy face' region of the embedding space.

@@ -533,8 +533,7 @@ def degenerate_embedding_test(
        f"(how tightly negatives cluster together)"
    )
    print(
-        f"  pos<->pos mean cos : {np.nanmean(pp):.3f}   "
-        f"(how tightly positives cluster)"
+        f"  pos<->pos mean cos : {np.nanmean(pp):.3f}   (how tightly positives cluster)"
    )
    print(
        f"  pos<->neg mean cos : {pn.mean():.3f}   "
@@ -558,11 +557,7 @@ def degenerate_embedding_test(
        neg_scores = np.array([cosine(n.embedding, clean_mean) for n in neg])
        neg_confs = np.array([similarity_to_confidence(c) for c in neg_scores])
        pos_scores = np.array(
-            [
-                cosine(pos[i].embedding, clean_mean)
-                for i in range(len(pos))
-                if keep[i]
-            ]
+            [cosine(pos[i].embedding, clean_mean) for i in range(len(pos)) if keep[i]]
        )
        print(
            f"\n  mean_intra >= {thresh}: keeping {int(keep.sum())}/{len(pos)} positives"
@@ -585,9 +580,7 @@ def degenerate_embedding_test(
        )


-def contamination_analysis(
-    pos: list[FaceSample], neg: list[FaceSample]
-) -> None:
+def contamination_analysis(pos: list[FaceSample], neg: list[FaceSample]) -> None:
    """Check whether the positive collection contains a second identity.

    Two signals:
@@ -617,10 +610,7 @@ def contamination_analysis(
        "\nPositives closer to a negative than to their own class avg"
        "\n(these are candidates for mislabeled images):"
    )
-    print(
-        f"\n{'max_neg':>7}  {'mean_neg':>8}  {'mean_intra':>10}  "
-        f"{'delta':>6}  name"
-    )
+    print(f"\n{'max_neg':>7}  {'mean_neg':>8}  {'mean_intra':>10}  {'delta':>6}  name")
    rows = list(zip(pos_names, max_to_neg, mean_to_neg, mean_intra))
    rows.sort(key=lambda r: -(r[1] - r[3]))
    for nm, mxn, mnn, mi in rows[:15]:
@@ -704,7 +694,9 @@ def main() -> int:
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog=__doc__,
    )
-    ap.add_argument("--positive", required=True, help="Training folder for one identity")
+    ap.add_argument(
+        "--positive", required=True, help="Training folder for one identity"
+    )
    ap.add_argument(
        "--negative",
        default=None,
@@ -0,0 +1,204 @@
+/**
+ * Camera live playback stream settings tests -- MEDIUM tier.
+ *
+ * The live streams field maps a display name to a go2rtc stream. Switching
+ * cameras from the selector keeps the form mounted and only swaps its data, so
+ * the stream name input has to follow the newly selected camera. It used to be
+ * an uncontrolled input, which left the previous camera's stream name on screen
+ * and renamed the wrong key if the stale text was ever committed.
+ *
+ * Renames are committed per keystroke so the section is marked as modified
+ * right away, except while the typed name belongs to another stream, since
+ * renaming onto an existing name merges the two entries.
+ */
+
+import { readFileSync } from "node:fs";
+import { resolve, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { test, expect } from "../../fixtures/frigate-test";
+import type { Page } from "@playwright/test";
+import { configFactory } from "../../fixtures/mock-data/config";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const CONFIG_SCHEMA = JSON.parse(
+  readFileSync(
+    resolve(__dirname, "../../fixtures/mock-data/config-schema.json"),
+    "utf-8",
+  ),
+);
+
+const GO2RTC_STREAMS = {
+  front_door_main: ["rtsp://user:pass@192.168.0.20:554/Stream1"],
+  backyard_main: ["rtsp://user:pass@192.168.0.21:554/Stream1"],
+};
+
+const CAMERA_LIVE_STREAMS = {
+  front_door: { front_door: "front_door_main" },
+  backyard: { backyard: "backyard_main" },
+};
+
+const SETTINGS_URL = "/settings?page=cameraLivePlayback&camera=front_door";
+
+async function installRoutes(
+  page: Page,
+  frontDoorStreams: Record<string, string> = CAMERA_LIVE_STREAMS.front_door,
+) {
+  const config = configFactory({
+    go2rtc: { streams: GO2RTC_STREAMS },
+    cameras: {
+      front_door: { live: { streams: frontDoorStreams } },
+      backyard: { live: { streams: CAMERA_LIVE_STREAMS.backyard } },
+    },
+  });
+
+  let lastSavedConfig: unknown = null;
+
+  await page.route("**/api/config/schema.json", (route) =>
+    route.fulfill({ json: CONFIG_SCHEMA }),
+  );
+  await page.route("**/api/config", (route) => {
+    if (route.request().method() === "GET") {
+      return route.fulfill({ json: config });
+    }
+    return route.fulfill({ json: { success: true } });
+  });
+  await page.route("**/api/config/raw_paths", (route) =>
+    route.fulfill({
+      json: {
+        go2rtc: { streams: GO2RTC_STREAMS },
+        cameras: {
+          front_door: { live: { streams: frontDoorStreams } },
+          backyard: { live: { streams: CAMERA_LIVE_STREAMS.backyard } },
+        },
+      },
+    }),
+  );
+  await page.route("**/api/config/set", async (route) => {
+    lastSavedConfig = route.request().postDataJSON();
+    await route.fulfill({ json: { success: true, require_restart: false } });
+  });
+
+  return { capturedConfig: () => lastSavedConfig };
+}
+
+async function selectCamera(page: Page, friendlyName: string) {
+  await page.getByRole("button", { name: "Select a camera" }).click();
+  await page.getByRole("switch", { name: friendlyName }).click();
+}
+
+function streamNameInputs(page: Page) {
+  return page.getByRole("textbox", { name: "Stream name" });
+}
+
+function streamNames(page: Page) {
+  return streamNameInputs(page).evaluateAll((inputs) =>
+    inputs.map((input) => (input as HTMLInputElement).value),
+  );
+}
+
+/** Rows render in config order, which is not the order they were declared in. */
+async function streamNameRow(page: Page, name: string) {
+  await expect.poll(() => streamNames(page)).toContain(name);
+  const names = await streamNames(page);
+  return streamNameInputs(page).nth(names.indexOf(name));
+}
+
+test.describe("camera live playback streams @medium", () => {
+  test("switching cameras updates the stream name field", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(frigateApp.page);
+    await frigateApp.goto(SETTINGS_URL);
+
+    const streamName = frigateApp.page.getByRole("textbox", {
+      name: "Stream name",
+    });
+    await expect(streamName).toHaveValue("front_door");
+    await expect(
+      frigateApp.page.getByRole("combobox", { name: "go2rtc stream" }),
+    ).toContainText("front_door_main");
+
+    await selectCamera(frigateApp.page, "Backyard");
+
+    await expect(streamName).toHaveValue("backyard");
+    await expect(
+      frigateApp.page.getByRole("combobox", { name: "go2rtc stream" }),
+    ).toContainText("backyard_main");
+  });
+
+  test("typing a new name enables Save without leaving the field", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(frigateApp.page);
+    await frigateApp.goto(SETTINGS_URL);
+
+    const save = frigateApp.page.getByRole("button", { name: "Save" });
+    await expect(save).toBeDisabled();
+
+    const streamName = await streamNameRow(frigateApp.page, "front_door");
+    await streamName.click();
+    await frigateApp.page.keyboard.press("End");
+    await frigateApp.page.keyboard.type("_hd");
+
+    // Still focused: the rename is committed per keystroke, not on blur.
+    await expect(save).toBeEnabled();
+    await expect(streamName).toBeFocused();
+    await expect(streamName).toHaveValue("front_door_hd");
+  });
+
+  test("typing through another stream's name keeps both streams", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(frigateApp.page, {
+      front: "front_door_main",
+      front_door: "backyard_main",
+    });
+    await frigateApp.goto(SETTINGS_URL);
+
+    const streamName = await streamNameRow(frigateApp.page, "front_door");
+    await streamName.click();
+    await frigateApp.page.keyboard.press("End");
+    // "front_door" passes through "front", which the other row already uses.
+    await frigateApp.page.keyboard.press("Backspace");
+    await frigateApp.page.keyboard.press("Backspace");
+    await frigateApp.page.keyboard.press("Backspace");
+    await frigateApp.page.keyboard.press("Backspace");
+    await frigateApp.page.keyboard.press("Backspace");
+    await expect(streamName).toHaveValue("front");
+    await frigateApp.page.keyboard.type("yard");
+    await streamName.blur();
+
+    expect(await streamNames(frigateApp.page)).toEqual(["frontyard", "front"]);
+  });
+
+  test("renaming a stream saves the new name for the selected camera", async ({
+    frigateApp,
+  }) => {
+    const capture = await installRoutes(frigateApp.page);
+    await frigateApp.goto(SETTINGS_URL);
+
+    await selectCamera(frigateApp.page, "Backyard");
+
+    const streamName = frigateApp.page.getByRole("textbox", {
+      name: "Stream name",
+    });
+    await expect(streamName).toHaveValue("backyard");
+    await streamName.fill("Backyard HD");
+    // The rename is committed on blur, not on every keystroke.
+    await streamName.blur();
+
+    await frigateApp.page.getByRole("button", { name: "Save" }).click();
+
+    await expect
+      .poll(() => capture.capturedConfig(), { timeout: 5_000 })
+      .toMatchObject({
+        config_data: {
+          cameras: {
+            backyard: {
+              live: { streams: { "Backyard HD": "backyard_main" } },
+            },
+          },
+        },
+      });
+  });
+});
@@ -0,0 +1,129 @@
+/**
+ * Semantic Search settings tests -- MEDIUM tier.
+ *
+ * Focuses on the model_size field, which is unused when a GenAI embeddings
+ * provider is selected as the semantic search model. The resolved config always
+ * reports model_size (it has a schema default of "small"), even when the YAML
+ * file has no such key. Clearing model_size for a provider used to run
+ * unconditionally, which falsely marked the section dirty on load and asked the
+ * backend to delete a key that wasn't in the config file (KeyError: 'model_size').
+ */
+
+import { readFileSync } from "node:fs";
+import { resolve, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { test, expect } from "../../fixtures/frigate-test";
+import type { Page } from "@playwright/test";
+import { configFactory } from "../../fixtures/mock-data/config";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const CONFIG_SCHEMA = JSON.parse(
+  readFileSync(
+    resolve(__dirname, "../../fixtures/mock-data/config-schema.json"),
+    "utf-8",
+  ),
+);
+
+const PROVIDER = "llama_cpp";
+const SETTINGS_URL = "/settings?page=integrationSemanticSearch";
+const NOT_APPLICABLE = "Not applicable for GenAI providers";
+const UNSAVED = "You have unsaved changes";
+
+type SemanticSearch = {
+  enabled?: boolean;
+  model?: string;
+  model_size?: string;
+};
+
+async function installRoutes(page: Page, semanticSearch: SemanticSearch) {
+  const config = configFactory({
+    genai: { [PROVIDER]: { provider: PROVIDER, roles: ["embeddings"] } },
+    semantic_search: semanticSearch,
+  });
+
+  let lastSavedConfig: unknown = null;
+
+  await page.route("**/api/config/schema.json", (route) =>
+    route.fulfill({ json: CONFIG_SCHEMA }),
+  );
+  await page.route("**/api/config", (route) => {
+    if (route.request().method() === "GET") {
+      return route.fulfill({ json: config });
+    }
+    return route.fulfill({ json: { success: true } });
+  });
+  await page.route("**/api/config/set", async (route) => {
+    lastSavedConfig = route.request().postDataJSON();
+    await route.fulfill({ json: { success: true, require_restart: false } });
+  });
+  await page.route("**/api/config/raw_paths", (route) =>
+    route.fulfill({ json: { semantic_search: semanticSearch } }),
+  );
+
+  return { capturedConfig: () => lastSavedConfig };
+}
+
+test.describe("semantic search model_size @medium", () => {
+  test("a provider with a defaulted model_size is not dirty on load", async ({
+    frigateApp,
+  }) => {
+    // model_size stays at its schema default ("small"), i.e. it is not present
+    // in the YAML. This mirrors the reported bug: selecting a GenAI provider and
+    // returning to the page.
+    await installRoutes(frigateApp.page, {
+      enabled: true,
+      model: PROVIDER,
+    });
+    await frigateApp.goto(SETTINGS_URL);
+
+    // The provider path is active: model_size shows "Not applicable".
+    await expect(frigateApp.page.getByText(NOT_APPLICABLE)).toBeVisible();
+
+    // Give any clearing effect time to fire, then confirm the section stayed
+    // clean (no phantom unsaved-changes banner, Save disabled).
+    await frigateApp.page.waitForTimeout(1000);
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeHidden();
+    await expect(
+      frigateApp.page.getByRole("button", { name: "Save", exact: true }),
+    ).toBeDisabled();
+  });
+
+  test("switching from a configured non-default model_size clears it", async ({
+    frigateApp,
+  }) => {
+    // A genuinely configured non-default model_size ("large") can only come from
+    // the YAML, so switching to a provider must still remove it.
+    const capture = await installRoutes(frigateApp.page, {
+      enabled: true,
+      model: "jinav2",
+      model_size: "large",
+    });
+    await frigateApp.goto(SETTINGS_URL);
+
+    // Starts clean on a Jina model.
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeHidden();
+
+    // Switch the model to the GenAI provider.
+    await frigateApp.page
+      .getByRole("combobox", { name: /Semantic search model/ })
+      .click();
+    await frigateApp.page.getByRole("option", { name: PROVIDER }).click();
+
+    // The change is now dirty and model_size is no longer applicable.
+    await expect(frigateApp.page.getByText(NOT_APPLICABLE)).toBeVisible();
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeVisible();
+
+    await frigateApp.page
+      .getByRole("button", { name: "Save", exact: true })
+      .click();
+
+    // The saved payload removes model_size (empty string = "remove" key).
+    await expect
+      .poll(() => capture.capturedConfig(), { timeout: 5_000 })
+      .toMatchObject({
+        config_data: {
+          semantic_search: { model: PROVIDER, model_size: "" },
+        },
+      });
+  });
+});
@@ -73,7 +73,7 @@
        "react-markdown": "^9.0.1",
        "react-router-dom": "^6.30.3",
        "react-swipeable": "^7.0.2",
-        "react-zoom-pan-pinch": "^3.7.0",
+        "react-zoom-pan-pinch": "3.4.4",
        "remark-gfm": "^4.0.0",
        "scroll-into-view-if-needed": "^3.1.0",
        "sonner": "^2.0.7",
@@ -12354,9 +12354,9 @@
      }
    },
    "node_modules/react-zoom-pan-pinch": {
-      "version": "3.7.0",
-      "resolved": "https://registry.npmjs.org/react-zoom-pan-pinch/-/react-zoom-pan-pinch-3.7.0.tgz",
-      "integrity": "sha512-UmReVZ0TxlKzxSbYiAj+LeGRW8s8LraAFTXRAxzMYnNRgGPsxCudwZKVkjvGmjtx7SW/hZamt69NUmGf4xrkXA==",
+      "version": "3.4.4",
+      "resolved": "https://registry.npmjs.org/react-zoom-pan-pinch/-/react-zoom-pan-pinch-3.4.4.tgz",
+      "integrity": "sha512-lGTu7D9lQpYEQ6sH+NSlLA7gicgKRW8j+D/4HO1AbSV2POvKRFzdWQ8eI0r3xmOsl4dYQcY+teV6MhULeg1xBw==",
      "license": "MIT",
      "engines": {
        "node": ">=8",
@@ -87,7 +87,7 @@
    "react-markdown": "^9.0.1",
    "react-router-dom": "^6.30.3",
    "react-swipeable": "^7.0.2",
-    "react-zoom-pan-pinch": "^3.7.0",
+    "react-zoom-pan-pinch": "3.4.4",
    "remark-gfm": "^4.0.0",
    "scroll-into-view-if-needed": "^3.1.0",
    "sonner": "^2.0.7",
@@ -61,7 +61,8 @@
      "error": {
        "endTimeMustAfterStartTime": "L'hora de finalització ha de ser posterior a l'hora d'inici",
        "noVaildTimeSelected": "No s'ha seleccionat un rang de temps vàlid",
-        "failed": "No s'ha pogut inciar l'exportació: {{error}}"
+        "failed": "No s'ha pogut inciar l'exportació: {{error}}",
+        "noValidTimeSelected": "No s'ha seleccionat cap interval de temps vàlid"
      },
      "view": "Vista",
      "queued": "Exporta a la cua. Mostra el progrés a la pàgina d'exportacions.",
@@ -493,6 +493,9 @@
            "max_concurrent": {
                "label": "Màxim d'exportacions concurrents",
                "description": "Nombre màxim de treballs d'exportació a processar al mateix temps."
+            },
+            "chapters": {
+                "label": "Metadades de capítol per incrustar en els enregistraments exportats"
            }
        },
        "preview": {
@@ -864,8 +867,8 @@
            "description": "Ordre numèric utilitzat per ordenar la càmera a la interfície d'usuari (taulell de control i llistes per defecte); els nombres més grans apareixen més tard."
        },
        "dashboard": {
-            "label": "Mostra a l'interfície d'usuari",
-            "description": "Estableix si aquesta càmera és visible a tot arreu a la interfície d'usuari de la Frigate. Desactivar això requerirà editar manualment la configuració per tornar a veure aquesta càmera a la interfície d'usuari."
+            "label": "Mostra al tauler en directe",
+            "description": "Alterna si aquesta càmera és visible al tauler de control en directe de totes les càmeres per defecte. La càmera roman disponible a tot arreu a la interfície d'usuari, inclosos els grups i la configuració de la càmera."
        },
        "review": {
            "label": "Mostra en la revisió",
@@ -380,6 +380,9 @@
            "max_concurrent": {
                "label": "Màxim d'exportacions concurrents",
                "description": "Nombre màxim de treballs d'exportació a processar al mateix temps."
+            },
+            "chapters": {
+                "label": "Metadades de capítol per incrustar en els enregistraments exportats"
            }
        },
        "preview": {
@@ -1917,7 +1920,7 @@
            },
            "model_type": {
                "label": "Tipus de Model de detecció d'objecte",
-                "description": "Tipus d'arquitectura del model de detector (ssd, yolox, yolonas) usat per l'optimització d'alguns detectors."
+                "description": "Tipus d'arquitectura del model de detector (ssd, yolox, yolonas, yolo-generic, rfdetr, dfine) utilitzat per alguns detectors per a l'optimització"
            }
        },
        "model_path": {
@@ -1966,7 +1969,7 @@
        },
        "model_type": {
            "label": "Tipus de model de detecció d'objectes",
-            "description": "Tipus d'arquitectura del model de detector (ssd, yolox, yolonas) utilitzat per alguns detectors per a l'optimització."
+            "description": "Tipus d'arquitectura del model de detector (ssd, yolox, yolonas, yolo-generic, rfdetr, dfine) utilitzat per alguns detectors per a l'optimització."
        }
    },
    "genai": {
@@ -2335,8 +2338,8 @@
            "description": "Ordre numèric utilitzat per ordenar la càmera a la interfície d'usuari (taulell de control i llistes per defecte); els nombres més grans apareixen més tard."
        },
        "dashboard": {
-            "label": "Mostra a la interfície",
-            "description": "Estableix si aquesta càmera és visible a tot arreu a la interfície d'usuari de Frigate. Desactivar això requerirà editar manualment la configuració per tornar a veure aquesta càmera a la interfície d'usuari."
+            "label": "Mostra al tauler en directe",
+            "description": "Alterna si aquesta càmera és visible al tauler de control en directe de totes les càmeres per defecte. La càmera roman disponible a tot arreu a la interfície d'usuari, inclosos els grups i la configuració de la càmera."
        },
        "review": {
            "label": "Mostra en la revisió",
@@ -126,5 +126,7 @@
  "baby_stroller": "Cotxet",
  "rickshaw": "Ricksaw",
  "Rodent": "Rosegador",
-  "rodent": "Rosegador"
+  "rodent": "Rosegador",
+  "possum": "Possum",
+  "garbage_truck": "Camió de brossa"
 }
@@ -192,7 +192,20 @@
    "title": "Edita el model de classificació",
    "descriptionState": "Edita les classes per a aquest model de classificació d'estats. Els canvis requeriran tornar a entrenar el model.",
    "descriptionObject": "Edita el tipus d'objecte i el tipus de classificació per a aquest model de classificació d'objectes.",
-    "stateClassesInfo": "Nota: Canviar les classes d'estat requereix tornar a entrenar el model amb les classes actualitzades."
+    "stateClassesInfo": "S'ha actualitzat el model. Restringeix el model perquè els canvis de classe tinguin efecte.",
+    "enabled": "Habilitat",
+    "enabledDesc": "Executa aquest model. Quan està desactivat, deixa d'executar-se i ja no classifica.",
+    "saveAttempts": "Desa els intents",
+    "saveAttemptsDesc": "Nombre d'imatges de classificació que s'intenten mantenir per a les classificacions recents UI.",
+    "motion": "Executa en moviment",
+    "motionDesc": "Executa la classificació quan es detecta el moviment dins de l'escapçat configurat.",
+    "interval": "Interval",
+    "intervalDesc": "Segons entre les classificacions periòdiques. Deixeu-ho buit per a executar-se només en moviment.",
+    "intervalPlaceholder": "Sense interval",
+    "errors": {
+      "saveAttemptsInvalid": "Els intents de desar han de ser un nombre sencer de 0 o més",
+      "intervalInvalid": "L'interval ha de ser un nombre sencer més gran que 0"
+    }
  },
  "tooltip": {
    "trainingInProgress": "El model s'està entrenant actualment",
@@ -202,5 +215,6 @@
  },
  "none": "Cap",
  "reclassifyImageAs": "Reclassifica la imatge com a:",
-  "reclassifyImage": "Reclassifica la imatge"
+  "reclassifyImage": "Reclassifica la imatge",
+  "disabled": "Desactivat"
 }
--- a/Show More
+++ b/Show More