Ollama image improvements (#24605)

* Support embeddinggemma for ollama

* Refactor docs

* Allow Ollama to probe cost of image tokens

* Cleanup

docs/docs/configuration/audio_detectors.md

@@ -289,7 +289,7 @@ The only field that is valid at the camera level is `enabled`. In particular `mo
 
 #### GenAI Provider
 
-Frigate can send audio to a GenAI provider for transcription when that provider has the `transcribe` role. This is useful if you already run a GenAI provider, or if you do not have the CPU/GPU headroom for a local whisper model. Supported providers are **OpenAI**, **Azure OpenAI**, **Gemini**, and **llama.cpp** with an audio-capable model (a dedicated ASR model such as Qwen3-ASR, or a general multimodal model that accepts audio). Ollama is not supported as it has no audio input.
+Frigate can send audio to a GenAI provider for transcription when that provider has the `transcribe` role. This is useful if you already run a GenAI provider, or if you do not have the CPU/GPU headroom for a local whisper model. See [Provider support](/configuration/genai/genai_config#provider-support) for which providers can serve this role. The model must accept audio: either a dedicated ASR model such as Qwen3-ASR, or a general multimodal model that accepts audio.
 
 To use a GenAI provider for audio transcription:
 

docs/docs/configuration/genai/config.md

@@ -43,7 +43,19 @@ genai:
 
 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`, `embeddings`, and `transcribe`. A provider handles the first three by default; `transcribe` must always be listed explicitly, and is not available on Ollama, which has no audio input. 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.
+Each provider handles one or more **roles**: `chat`, `descriptions`, `embeddings`, and `transcribe`. A provider handles the first three by default; `transcribe` must always be listed explicitly. 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. Not every provider supports every role; see [Provider support](#provider-support).
+
+### Provider support
+
+| Provider                      | Descriptions | Chat | Embeddings | Transcription |
+| ----------------------------- | :----------: | :--: | :--------: | :-----------: |
+| llama.cpp (`llamacpp`)        |      ✅      |  ✅  |     ✅     |      ✅       |
+| Ollama (`ollama`)             |      ✅      |  ✅  |     ✅     |      ❌       |
+| OpenAI (`openai`)             |      ✅      |  ✅  |     ❌     |      ✅       |
+| Azure OpenAI (`azure_openai`) |      ✅      |  ✅  |     ❌     |      ✅       |
+| Google Gemini (`gemini`)      |      ✅      |  ✅  |     ❌     |      ✅       |
+
+A ✅ means Frigate can use the provider for that feature. The configured model must also support it: a vision model for descriptions and chat, a multimodal embedding model for embeddings (see [Embedding models](#embedding-models)), and an audio-capable model for transcription. Some features also need extra provider setup, covered in each provider's section below. OpenAI-compatible servers use the `openai` provider, so they follow the OpenAI row.
 
 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_`.
 
@@ -73,9 +85,10 @@ You must use a vision-capable model with Frigate. The following models are recom
 
 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`. |
+| Model                | Notes                                                                                                                                                                  |
+| -------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `embeddinggemma-2`   | Multimodal embeddings for [Semantic Search](/configuration/semantic_search#genai-provider). Strong semantic search accuracy with efficient inference on a small model. |
+| `qwen3-vl-embedding` | Multimodal embeddings for [Semantic Search](/configuration/semantic_search#genai-provider). Good performance, large model that requires strong hardware for inference. |
 
 #### Transcription models
 
@@ -152,6 +165,10 @@ genai:
 
 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.
 
+#### Embeddings
+
+To serve the `embeddings` role for [Semantic Search](/configuration/semantic_search#genai-provider), start the llama.cpp server with `--embeddings`, plus `--mmproj` for image support. See the [llama.cpp server documentation](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md) for details.
+
 ### 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.
@@ -197,6 +214,10 @@ genai:
 </TabItem>
 </ConfigTabs>
 
+#### Embeddings
+
+Ollama can serve the `embeddings` role for [Semantic Search](/configuration/semantic_search#genai-provider). Embedding images requires Ollama 0.40.1 or newer and an embedding model with a vision encoder, such as `embeddinggemma-2:440m`. For a saved provider, the UI hides the role unless Ollama reports its model as an embedding model. Use a separate provider entry for the embedding model rather than adding the role to a vision chat model.
+
 ### OpenAI-Compatible
 
 Frigate supports any provider that implements the OpenAI API standard. This includes self-hosted solutions like [vLLM](https://docs.vllm.ai/), [LocalAI](https://localai.io/), and other OpenAI-compatible servers.

docs/docs/configuration/semantic_search.md

@@ -133,13 +133,12 @@ Switching between V1 and V2 requires reindexing your embeddings. The embeddings
 
 ### GenAI Provider
 
-Frigate can use a GenAI provider for semantic search embeddings when that provider has the `embeddings` role. Currently, only **llama.cpp** supports multimodal embeddings (both text and images).
+Frigate can use a GenAI provider for semantic search embeddings when that provider has the `embeddings` role. See [Provider support](/configuration/genai/genai_config#provider-support) for which providers can serve this role.
 
-To use llama.cpp for semantic search:
+To use a GenAI provider for semantic search:
 
-1. Configure a GenAI provider with `embeddings` in its `roles`.
+1. Configure a GenAI provider with `embeddings` in its `roles`, using a multimodal embedding model (both text and images). See [Embedding models](/configuration/genai/genai_config#embedding-models) for recommendations, and your provider's section of the [GenAI docs](/configuration/genai/genai_config) for any extra setup it needs.
 2. Set the semantic search model to the GenAI config key (e.g. `default`).
-3. Start the llama.cpp server with `--embeddings` and `--mmproj` for image support.
 
 <ConfigTabs>
 <TabItem value="ui">
@@ -174,8 +173,6 @@ semantic_search:
 </TabItem>
 </ConfigTabs>
 
-The llama.cpp server must be started with `--embeddings` for the embeddings API, and a multi-modal embeddings model. See the [llama.cpp server documentation](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md) for details.
-
 :::note
 
 Switching between Jina models and a GenAI provider requires reindexing. Embeddings from different backends are incompatible.

frigate/genai/__init__.py

@@ -22,6 +22,7 @@ from frigate.genai.prompts import (
     build_review_description_response_format,
     build_review_summary_prompt,
 )
+from frigate.genai.utils import synthetic_jpeg
 from frigate.models import Event
 from frigate.util.builtin import has_non_finite_number
 
@@ -66,6 +67,8 @@ class GenAIClient:
         self.genai_config: GenAIConfig = genai_config
         self.timeout = timeout
         self.validate_model = validate_model
+        self._image_token_cache: dict[tuple[int, int], int] = {}
+        self._text_baseline_tokens: int | None = None
         self.provider = self._init_provider()
         self._last_init_attempt = time.monotonic()
 
@@ -372,10 +375,63 @@ class GenAIClient:
     def estimate_image_tokens(self, width: int, height: int) -> float:
         """Estimate prompt tokens consumed by a single image of the given dimensions.
 
-        Default heuristic: ~1 token per 1250 pixels. Providers that can measure or
-        know their model's exact image-token cost should override.
+        Providers that implement ``_count_prompt_tokens`` are probed for the
+        model's real cost: the same minimal prompt is counted with and without a
+        synthetic image, and the difference is cached per (width, height) since
+        image tokenization depends only on the dimensions and the loaded model.
+        Otherwise, or if probing fails, falls back to ~1 token per 1250 pixels.
         """
-        return (width * height) / 1250
+        heuristic = (width * height) / 1250
+
+        if self.provider is None:
+            return heuristic
+
+        cached = self._image_token_cache.get((width, height))
+
+        if cached is not None:
+            return cached
+
+        try:
+            if self._text_baseline_tokens is None:
+                self._text_baseline_tokens = self._count_prompt_tokens(None)
+
+            if self._text_baseline_tokens is None:
+                return heuristic
+
+            with_image = self._count_prompt_tokens(synthetic_jpeg(width, height))
+        except Exception as e:
+            logger.debug(
+                "%s image-token probe failed for %dx%d (%s); using heuristic",
+                self.__class__.__name__,
+                width,
+                height,
+                e,
+            )
+            return heuristic
+
+        if with_image is None:
+            return heuristic
+
+        tokens = max(1, with_image - self._text_baseline_tokens)
+        self._image_token_cache[(width, height)] = tokens
+        logger.debug(
+            "%s model '%s' uses ~%d tokens for %dx%d images",
+            self.__class__.__name__,
+            self.genai_config.model,
+            tokens,
+            width,
+            height,
+        )
+        return tokens
+
+    def _count_prompt_tokens(self, image: bytes | None) -> int | None:
+        """Prompt tokens the provider reports for a minimal "." request, with
+        ``image`` attached when given.
+
+        Return None when the provider cannot report prompt tokens; raise on
+        request failures. Used by estimate_image_tokens.
+        """
+        return None
 
     def embed(
         self,

frigate/genai/plugins/llama_cpp.py

@@ -1,7 +1,6 @@
 """llama.cpp Provider for Frigate AI."""
 
 import base64
-import io
 import json
 import logging
 from collections.abc import AsyncGenerator
@@ -10,11 +9,14 @@ from typing import Any, cast
 import httpx
 import numpy as np
 import requests
-from PIL import Image
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
-from frigate.genai.utils import interleave_images, parse_tool_calls_from_message
+from frigate.genai.utils import (
+    interleave_images,
+    parse_tool_calls_from_message,
+    to_jpeg,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -76,20 +78,6 @@ def _parse_launch_arg(args: list[str], flag: str) -> str | None:
     return args[idx + 1]
 
 
-def _to_jpeg(img_bytes: bytes) -> bytes | None:
-    """Convert image bytes to JPEG. llama.cpp/STB does not support WebP."""
-    try:
-        img = Image.open(io.BytesIO(img_bytes))
-        if img.mode != "RGB":
-            img = img.convert("RGB")  # type: ignore[assignment]
-        buf = io.BytesIO()
-        img.save(buf, format="JPEG", quality=85)
-        return buf.getvalue()
-    except Exception as e:
-        logger.warning("Failed to convert image to JPEG: %s", e)
-        return None
-
-
 @register_genai_provider(GenAIProviderEnum.llamacpp)
 class LlamaCppClient(GenAIClient):
     """Generative AI client for Frigate using llama.cpp server."""
@@ -101,8 +89,6 @@ class LlamaCppClient(GenAIClient):
     _supports_audio: bool
     _supports_tools: bool
     _supports_reasoning: bool
-    _image_token_cache: dict[tuple[int, int], int]
-    _text_baseline_tokens: int | None
 
     @property
     def supports_embeddings(self) -> bool:
@@ -156,8 +142,6 @@ class LlamaCppClient(GenAIClient):
         self._supports_audio = False
         self._supports_tools = False
         self._supports_reasoning = False
-        self._image_token_cache = {}
-        self._text_baseline_tokens = None
 
         base_url = (
             self.genai_config.base_url.rstrip("/")
@@ -601,78 +585,23 @@ class LlamaCppClient(GenAIClient):
             return self._context_size
         return 4096
 
-    def estimate_image_tokens(self, width: int, height: int) -> float:
-        """Probe the llama.cpp server to learn the model's image-token cost at the
-        requested dimensions.
-
-        llama.cpp's image tokenization is a deterministic function of dimensions and
-        the loaded mmproj, so the result is cached per (width, height) for the
-        lifetime of the process. Falls back to the base pixel heuristic if the
-        server is unreachable or the response is malformed.
-        """
-        if self.provider is None:
-            return super().estimate_image_tokens(width, height)
-
-        cached = self._image_token_cache.get((width, height))
-
-        if cached is not None:
-            return cached
-
-        try:
-            baseline = self._probe_baseline_tokens()
-            with_image = self._probe_image_prompt_tokens(width, height)
-            tokens = max(1, with_image - baseline)
-        except Exception as e:
-            logger.debug(
-                "llama.cpp image-token probe failed for %dx%d (%s); using heuristic",
-                width,
-                height,
-                e,
-            )
-            return super().estimate_image_tokens(width, height)
-
-        self._image_token_cache[(width, height)] = tokens
-        logger.debug(
-            "llama.cpp model '%s' uses ~%d tokens for %dx%d images",
-            self.genai_config.model,
-            tokens,
-            width,
-            height,
-        )
-        return tokens
-
-    def _probe_baseline_tokens(self) -> int:
-        """Return prompt_tokens for a minimal text-only request. Cached after first call."""
-        if self._text_baseline_tokens is not None:
-            return self._text_baseline_tokens
-
-        self._text_baseline_tokens = self._probe_prompt_tokens(
-            [{"type": "text", "text": "."}]
-        )
-        return self._text_baseline_tokens
-
-    def _probe_image_prompt_tokens(self, width: int, height: int) -> int:
-        """Return prompt_tokens for a single synthetic image plus minimal text."""
-        img = Image.new("RGB", (width, height), (128, 128, 128))
-        buf = io.BytesIO()
-        img.save(buf, format="JPEG", quality=60)
-        encoded = base64.b64encode(buf.getvalue()).decode("utf-8")
-        return self._probe_prompt_tokens(
-            [
-                {"type": "text", "text": "."},
-                {
-                    "type": "image_url",
-                    "image_url": {"url": f"data:image/jpeg;base64,{encoded}"},
-                },
-            ]
-        )
-
-    def _probe_prompt_tokens(self, content: list[dict[str, Any]]) -> int:
+    def _count_prompt_tokens(self, image: bytes | None) -> int | None:
         """POST a 1-token chat completion and return reported prompt_tokens.
 
         Uses a generous timeout to absorb a cold model load on the first probe
         when the server lazily loads models on demand (e.g. llama-swap).
         """
+        content: list[dict[str, Any]] = [{"type": "text", "text": "."}]
+
+        if image is not None:
+            encoded = base64.b64encode(image).decode("utf-8")
+            content.append(
+                {
+                    "type": "image_url",
+                    "image_url": {"url": f"data:image/jpeg;base64,{encoded}"},
+                }
+            )
+
         payload = {
             "model": self.genai_config.model,
             "messages": [{"role": "user", "content": content}],
@@ -807,7 +736,7 @@ class LlamaCppClient(GenAIClient):
 
         for img in images:
             # llama.cpp uses STB which does not support WebP; convert to JPEG
-            jpeg_bytes = _to_jpeg(img)
+            jpeg_bytes = to_jpeg(img)
             to_encode = jpeg_bytes if jpeg_bytes is not None else img
             encoded = base64.b64encode(to_encode).decode("utf-8")
             # The trailing newline keeps tokenization identical to the older

frigate/genai/plugins/ollama.py

@@ -7,6 +7,7 @@ import logging
 from collections.abc import AsyncGenerator
 from typing import Any
 
+import numpy as np
 from httpx import RemoteProtocolError, TimeoutException
 from ollama import AsyncClient as OllamaAsyncClient
 from ollama import Client as ApiClient
@@ -14,7 +15,11 @@ from ollama import ResponseError
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
-from frigate.genai.utils import interleave_images, parse_tool_calls_from_message
+from frigate.genai.utils import (
+    interleave_images,
+    parse_tool_calls_from_message,
+    to_jpeg,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -124,23 +129,39 @@ class OllamaClient(GenAIClient):
 
     provider: ApiClient | None
     provider_options: dict[str, Any]
+    _capabilities_cache: list[str] | None = None
     _supports_thinking_cache: bool | None = None
 
+    def _model_capabilities(self) -> list[str] | None:
+        """Capabilities Ollama reports for the configured model, or None when
+        they could not be fetched. Only successful lookups are cached."""
+        if self._capabilities_cache is not None:
+            return self._capabilities_cache
+        if self.provider is None:
+            return None
+        try:
+            response = self.provider.show(self.genai_config.model)
+        except Exception as e:
+            logger.debug("Failed to query Ollama model capabilities: %s", e)
+            return None
+        self._capabilities_cache = list(response.get("capabilities") or [])
+        return self._capabilities_cache
+
     @property
     def supports_toggleable_thinking(self) -> bool:
         if self._supports_thinking_cache is not None:
             return self._supports_thinking_cache
-        if self.provider is None:
+        capabilities = self._model_capabilities()
+        if capabilities is None:
             return False
-        try:
-            response = self.provider.show(self.genai_config.model)
-            capabilities = response.get("capabilities") or []
-            self._supports_thinking_cache = "thinking" in capabilities
-        except Exception as e:
-            logger.debug("Failed to query Ollama model capabilities: %s", e)
-            self._supports_thinking_cache = False
+        self._supports_thinking_cache = "thinking" in capabilities
         return self._supports_thinking_cache
 
+    @property
+    def supports_embeddings(self) -> bool:
+        """Whether Ollama reports the configured model as an embedding model."""
+        return "embedding" in (self._model_capabilities() or [])
+
     def _auth_headers(self) -> dict | None:
         if self.genai_config.api_key:
             return {"Authorization": "Bearer " + self.genai_config.api_key}
@@ -322,6 +343,93 @@ class OllamaClient(GenAIClient):
             self.genai_config.provider_options.get("options", {}).get("num_ctx", 4096)
         )
 
+    def _count_prompt_tokens(self, image: bytes | None) -> int | None:
+        """Send a 1-token chat request and return Ollama's prompt_eval_count.
+
+        Reuses the description request options so the probe runs with the same
+        num_ctx; a different value would make Ollama reload the model.
+        """
+        if self.provider is None:
+            return None
+
+        message: dict[str, Any] = {"role": "user", "content": "."}
+
+        if image is not None:
+            message["images"] = [image]
+
+        request_params = self._build_request_params(
+            [message], None, None, enable_thinking=False
+        )
+        request_params["options"] = {
+            **(request_params.get("options") or {}),
+            "num_predict": 1,
+        }
+        response = self.provider.chat(**request_params)
+        count = response.get("prompt_eval_count")
+        return int(count) if count is not None else None
+
+    def embed(
+        self,
+        texts: list[str] | None = None,
+        images: list[bytes] | None = None,
+    ) -> list[np.ndarray]:
+        """Generate embeddings via Ollama's /api/embed endpoint.
+
+        Each text is a plain string in `input` and each image is an
+        ``{"image": <base64>}`` item. Image input requires Ollama 0.40.1 or
+        newer and a model with a vision encoder (e.g. embeddinggemma-2:440m).
+        """
+        if self.provider is None:
+            logger.warning(
+                "Ollama provider has not been initialized. Check your Ollama configuration."
+            )
+            return []
+
+        texts = texts or []
+        images = images or []
+        if not texts and not images:
+            return []
+
+        inputs: list[str | dict[str, str]] = list(texts)
+        for img in images:
+            jpeg_bytes = to_jpeg(img)
+            to_encode = jpeg_bytes if jpeg_bytes is not None else img
+            inputs.append({"image": base64.b64encode(to_encode).decode("utf-8")})
+
+        payload: dict[str, Any] = {"model": self.genai_config.model, "input": inputs}
+        for key in ("options", "keep_alive"):
+            if key in self.genai_config.provider_options:
+                payload[key] = self.genai_config.provider_options[key]
+
+        try:
+            # The ollama SDK's embed() validates input as strings only, so
+            # image items have to bypass it and post the JSON directly.
+            response = self.provider._request_raw("POST", "/api/embed", json=payload)
+            body = response.json()
+        except (
+            TimeoutException,
+            ResponseError,
+            RemoteProtocolError,
+            ConnectionError,
+            ValueError,
+        ) as e:
+            logger.warning("Ollama embeddings error: %s", str(e))
+            return []
+
+        vectors = body.get("embeddings") if isinstance(body, dict) else None
+        if not isinstance(vectors, list):
+            logger.warning("Ollama embeddings returned unexpected format")
+            return []
+
+        if len(vectors) != len(inputs):
+            logger.warning(
+                "Ollama returned %d embeddings for %d inputs",
+                len(vectors),
+                len(inputs),
+            )
+
+        return [np.asarray(v, dtype=np.float32).flatten() for v in vectors]
+
     def _build_request_params(
         self,
         messages: list[dict[str, Any]],

frigate/genai/utils.py

@@ -1,12 +1,42 @@
 """Shared helpers for GenAI providers and chat (OpenAI-style messages, tool call parsing)."""
 
+import io
 import json
 import logging
 from typing import Any
 
+from PIL import Image
+
 logger = logging.getLogger(__name__)
 
 
+def to_jpeg(img_bytes: bytes) -> bytes | None:
+    """Convert image bytes to JPEG.
+
+    Some provider image decoders (e.g. llama.cpp's STB) do not support WebP,
+    which is the format Frigate stores thumbnails in.
+    """
+    try:
+        img = Image.open(io.BytesIO(img_bytes))
+        if img.mode != "RGB":
+            img = img.convert("RGB")  # type: ignore[assignment]
+        buf = io.BytesIO()
+        img.save(buf, format="JPEG", quality=85)
+        return buf.getvalue()
+    except Exception as e:
+        logger.warning("Failed to convert image to JPEG: %s", e)
+        return None
+
+
+def synthetic_jpeg(width: int, height: int) -> bytes:
+    """A flat gray JPEG of the given dimensions, for measuring image token cost."""
+    buf = io.BytesIO()
+    Image.new("RGB", (width, height), (128, 128, 128)).save(
+        buf, format="JPEG", quality=60
+    )
+    return buf.getvalue()
+
+
 def interleave_images(
     prompt: str, images: list[bytes], captions: list[str] | None = None
 ) -> list[str | bytes]:

frigate/test/test_genai_providers.py

@@ -427,6 +427,148 @@ class TestOllamaProvider(unittest.TestCase):
         self.assertEqual(message["content"], "prompt\n[img]")
         self.assertEqual(message["images"], [b"a"])
 
+    def test_capabilities_drive_embeddings_and_thinking(self):
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider.show.return_value = {"capabilities": ["embedding", "vision"]}
+
+        self.assertTrue(client.supports_embeddings)
+        self.assertFalse(client.supports_toggleable_thinking)
+        client.provider.show.assert_called_once()
+
+    def test_capability_lookup_failure_is_not_cached(self):
+        from ollama import ResponseError
+
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider.show.side_effect = [
+            ResponseError("unavailable", 503),
+            {"capabilities": ["embedding"]},
+        ]
+
+        self.assertFalse(client.supports_embeddings)
+        self.assertTrue(client.supports_embeddings)
+
+    def test_thinking_rechecked_after_provider_recovers(self):
+        client = self._client()
+        client.provider = None
+
+        self.assertFalse(client.supports_toggleable_thinking)
+
+        client.provider = MagicMock()
+        client.provider.show.return_value = {"capabilities": ["thinking"]}
+
+        self.assertTrue(client.supports_toggleable_thinking)
+        params = client._build_request_params(
+            [{"role": "user", "content": "hi"}], None, None, enable_thinking=True
+        )
+        self.assertTrue(params["think"])
+
+    @staticmethod
+    def _webp_bytes():
+        import io
+
+        from PIL import Image
+
+        buf = io.BytesIO()
+        Image.new("RGB", (8, 8), (200, 10, 10)).save(buf, format="WEBP")
+        return buf.getvalue()
+
+    def test_embed_posts_text_and_image_items(self):
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider._request_raw.return_value.json.return_value = {
+            "embeddings": [[0.1] * 768, [0.2] * 768]
+        }
+
+        result = client.embed(texts=["a person"], images=[self._webp_bytes()])
+
+        args = client.provider._request_raw.call_args
+        self.assertEqual(args.args, ("POST", "/api/embed"))
+        payload = args.kwargs["json"]
+        self.assertEqual(payload["model"], "llama3")
+        self.assertEqual(payload["input"][0], "a person")
+        self.assertEqual(list(payload["input"][1]), ["image"])
+        # WebP thumbnails are converted to JPEG before being sent
+        image = base64.b64decode(payload["input"][1]["image"])
+        self.assertEqual(image[:2], b"\xff\xd8")
+        self.assertEqual(len(result), 2)
+        self.assertAlmostEqual(float(result[1][0]), 0.2, places=5)
+
+    def test_embed_passes_configured_options(self):
+        client = _make_client(
+            "ollama",
+            model="embeddinggemma-2",
+            base_url="http://localhost:9999",
+            provider_options={"options": {"num_ctx": 2048}, "keep_alive": "10m"},
+        )
+        client.provider = MagicMock()
+        client.provider._request_raw.return_value.json.return_value = {
+            "embeddings": [[0.1] * 768]
+        }
+
+        client.embed(texts=["a"])
+
+        payload = client.provider._request_raw.call_args.kwargs["json"]
+        self.assertEqual(payload["options"], {"num_ctx": 2048})
+        self.assertEqual(payload["keep_alive"], "10m")
+
+    @staticmethod
+    def _chat_counting_prompt_tokens(**params):
+        """Fake chat that reports 10 prompt tokens plus 250 per image."""
+        images = params["messages"][0].get("images") or []
+        return {
+            "message": {"content": "."},
+            "prompt_eval_count": 10 + 250 * len(images),
+        }
+
+    def test_image_tokens_probed_with_one_token_requests(self):
+        client = _make_client(
+            "ollama",
+            model="qwen3-vl",
+            base_url="http://localhost:9999",
+            provider_options={"options": {"num_ctx": 16384}},
+        )
+        client.provider = MagicMock()
+        client.provider.chat.side_effect = self._chat_counting_prompt_tokens
+        client._supports_thinking_cache = False
+
+        self.assertEqual(client.estimate_image_tokens(320, 180), 250)
+
+        calls = client.provider.chat.call_args_list
+        self.assertEqual(len(calls), 2)
+        for call in calls:
+            self.assertEqual(
+                call.kwargs["options"], {"num_ctx": 16384, "num_predict": 1}
+            )
+        image = calls[1].kwargs["messages"][0]["images"][0]
+        self.assertEqual(image[:2], b"\xff\xd8")
+
+    def test_image_token_probe_error_uses_heuristic(self):
+        from ollama import ResponseError
+
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider.chat.side_effect = [
+            {"message": {"content": "."}, "prompt_eval_count": 10},
+            ResponseError("model does not support images", 400),
+        ]
+        client._supports_thinking_cache = False
+
+        self.assertEqual(client.estimate_image_tokens(250, 100), 20)
+        self.assertEqual(client._image_token_cache, {})
+
+    def test_embed_server_error_returns_empty(self):
+        from ollama import ResponseError
+
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider._request_raw.side_effect = ResponseError(
+            "model does not support media embeddings", 400
+        )
+
+        self.assertEqual(client.embed(images=[self._webp_bytes()]), [])
+
 
 # ---------------------------------------------------------------------------
 # llama.cpp
@@ -638,6 +780,46 @@ class TestLlamaCppProvider(unittest.TestCase):
         self.assertEqual([r.shape for r in result], [(768,), (768,)])
         self.assertEqual(float(result[1][-1]), 0.0)
 
+    @staticmethod
+    def _post_counting_prompt_tokens(url, json=None, timeout=None):
+        """Fake chat completion: 12 prompt tokens plus 300 per image part."""
+        content = json["messages"][0]["content"]
+        images = [p for p in content if p["type"] == "image_url"]
+        response = MagicMock()
+        response.json.return_value = {
+            "usage": {"prompt_tokens": 12 + 300 * len(images)}
+        }
+        return response
+
+    def test_image_tokens_probed_once_per_dimension(self):
+        client = self._client()
+
+        with patch.object(
+            client, "_post", side_effect=self._post_counting_prompt_tokens
+        ) as post:
+            self.assertEqual(client.estimate_image_tokens(320, 180), 300)
+            self.assertEqual(client.estimate_image_tokens(320, 180), 300)
+            self.assertEqual(client.estimate_image_tokens(640, 360), 300)
+
+        # one shared text baseline, then one image request per new dimension
+        self.assertEqual(post.call_count, 3)
+        payload = post.call_args_list[0].kwargs["json"]
+        self.assertEqual(payload["max_tokens"], 1)
+        self.assertEqual(
+            post.call_args_list[0].args[0], "http://localhost:9999/v1/chat/completions"
+        )
+
+    def test_image_token_probe_failure_is_not_cached(self):
+        client = self._client()
+
+        with patch.object(
+            client, "_post", side_effect=requests.exceptions.ConnectionError("down")
+        ):
+            self.assertEqual(client.estimate_image_tokens(250, 100), 20)
+
+        self.assertEqual(client._image_token_cache, {})
+        self.assertIsNone(client._text_baseline_tokens)
+
     def test_embed_request_error_returns_empty(self):
         client = self._client()
         response = MagicMock()
@@ -903,6 +1085,14 @@ class TestLlamaCppTranscribe(unittest.TestCase):
         post.assert_not_called()
 
 
+class TestImageTokenEstimate(unittest.TestCase):
+    def test_provider_without_token_counts_uses_heuristic(self):
+        client = _make_client("gemini", model="m", api_key="k")
+
+        self.assertEqual(client.estimate_image_tokens(250, 100), 20)
+        self.assertEqual(client._image_token_cache, {})
+
+
 class TestBaseClientTranscribe(unittest.TestCase):
     """Providers that don't implement the role must be inert, not broken."""