Improve live context image handling

frigate/api/chat.py

@@ -201,10 +201,9 @@ def get_tool_definitions() -> List[Dict[str, Any]]:
             "function": {
                 "name": "get_live_context",
                 "description": (
-                    "Get the current detection information for a camera: objects being tracked, "
+                    "Get the current live image and detection information for a camera: objects being tracked, "
                     "zones, timestamps. Use this to understand what is visible in the live view. "
-                    "Call this when the user has included a live image (via include_live_image) or "
+                    "Call this when answering questions about what is happening right now on a specific camera."
-                    "when answering questions about what is happening right now on a specific camera."
                 ),
                 "parameters": {
                     "type": "object",
@@ -384,12 +383,34 @@ async def _execute_get_live_context(
                     "stationary": obj_dict.get("stationary", False),
                 }
 
-        return {
+        result: Dict[str, Any] = {
             "camera": camera,
             "timestamp": frame_time,
             "detections": list(tracked_objects_dict.values()),
         }
 
+        # Grab live frame and handle based on provider configuration
+        image_url = await _get_live_frame_image_url(request, camera, allowed_cameras)
+        if image_url:
+            genai_manager = request.app.genai_manager
+            if genai_manager.tool_client is genai_manager.vision_client:
+                # Same provider handles both roles — pass image URL so it can
+                # be injected as a user message (images can't be in tool results)
+                result["_image_url"] = image_url
+            elif genai_manager.vision_client is not None:
+                # Separate vision provider — have it describe the image
+                frame_bytes = _decode_data_url(image_url)
+                if frame_bytes:
+                    description = genai_manager.vision_client._send(
+                        "Describe what you see in this security camera image. "
+                        "Focus on people, vehicles, animals, and any notable activity.",
+                        [frame_bytes],
+                    )
+                    if description:
+                        result["image_description"] = description
+
+        return result
+
     except Exception as e:
         logger.error(f"Error executing get_live_context: {e}", exc_info=True)
         return {
@@ -437,6 +458,17 @@ async def _get_live_frame_image_url(
         return None
 
 
+def _decode_data_url(data_url: str) -> Optional[bytes]:
+    """Decode a base64 data URL to raw bytes."""
+    try:
+        # Format: data:image/jpeg;base64,<data>
+        _, encoded = data_url.split(",", 1)
+        return base64.b64decode(encoded)
+    except (ValueError, Exception) as e:
+        logger.debug("Failed to decode data URL: %s", e)
+        return None
+
+
 async def _execute_set_camera_state(
     request: Request,
     arguments: Dict[str, Any],
@@ -527,12 +559,18 @@ async def _execute_pending_tools(
     pending_tool_calls: List[Dict[str, Any]],
     request: Request,
     allowed_cameras: List[str],
-) -> tuple[List[ToolCall], List[Dict[str, Any]]]:
+) -> tuple[List[ToolCall], List[Dict[str, Any]], List[Dict[str, Any]]]:
     """
-    Execute a list of tool calls; return (ToolCall list for API response, tool result dicts for conversation).
+    Execute a list of tool calls.
+
+    Returns:
+        (ToolCall list for API response,
+         tool result dicts for conversation,
+         extra messages to inject after tool results — e.g. user messages with images)
     """
     tool_calls_out: List[ToolCall] = []
     tool_results: List[Dict[str, Any]] = []
+    extra_messages: List[Dict[str, Any]] = []
     for tool_call in pending_tool_calls:
         tool_name = tool_call["name"]
         tool_args = tool_call.get("arguments") or {}
@@ -569,6 +607,27 @@ async def _execute_pending_tools(
                     for evt in tool_result
                     if isinstance(evt, dict)
                 ]
+
+            # Extract _image_url from get_live_context results — images can
+            # only be sent in user messages, not tool results
+            if isinstance(tool_result, dict) and "_image_url" in tool_result:
+                image_url = tool_result.pop("_image_url")
+                extra_messages.append(
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "text",
+                                "text": f"Here is the current live image from camera '{tool_result.get('camera', 'unknown')}'.",
+                            },
+                            {
+                                "type": "image_url",
+                                "image_url": {"url": image_url},
+                            },
+                        ],
+                    }
+                )
+
             result_content = (
                 json.dumps(tool_result)
                 if isinstance(tool_result, (dict, list))
@@ -604,7 +663,7 @@ async def _execute_pending_tools(
                     "content": error_content,
                 }
             )
-    return (tool_calls_out, tool_results)
+    return (tool_calls_out, tool_results, extra_messages)
 
 
 @router.post(
@@ -670,14 +729,6 @@ async def chat_completion(
             + "\n\nWhen users refer to cameras by their friendly name (e.g., 'Back Deck Camera'), use the corresponding camera ID (e.g., 'back_deck_cam') in tool calls."
         )
 
-    live_image_note = ""
-    if body.include_live_image:
-        live_image_note = (
-            f"\n\nThe first user message includes a live image from camera "
-            f"'{body.include_live_image}'. Use get_live_context for that camera to get "
-            "current detection details (objects, zones) to aid in understanding the image."
-        )
-
     system_prompt = f"""You are a helpful assistant for Frigate, a security camera NVR system. You help users answer questions about their cameras, detected objects, and events.
 
 Current server local date and time: {current_date_str} at {current_time_str}
@@ -687,7 +738,7 @@ Do not start your response with phrases like "I will check...", "Let me see...",
 Always present times to the user in the server's local timezone. When tool results include start_time_local and end_time_local, use those exact strings when listing or describing detection times—do not convert or invent timestamps. Do not use UTC or ISO format with Z for the user-facing answer unless the tool result only provides Unix timestamps without local time fields.
 When users ask about "today", "yesterday", "this week", etc., use the current date above as reference.
 When searching for objects or events, use ISO 8601 format for dates (e.g., {current_date_str}T00:00:00Z for the start of today).
-Always be accurate with time calculations based on the current date provided.{cameras_section}{live_image_note}"""
+Always be accurate with time calculations based on the current date provided.{cameras_section}"""
 
     conversation.append(
         {
@@ -696,7 +747,6 @@ Always be accurate with time calculations based on the current date provided.{ca
         }
     )
 
-    first_user_message_seen = False
     for msg in body.messages:
         msg_dict = {
             "role": msg.role,
@@ -707,21 +757,6 @@ Always be accurate with time calculations based on the current date provided.{ca
         if msg.name:
             msg_dict["name"] = msg.name
 
-        if (
-            msg.role == "user"
-            and not first_user_message_seen
-            and body.include_live_image
-        ):
-            first_user_message_seen = True
-            image_url = await _get_live_frame_image_url(
-                request, body.include_live_image, allowed_cameras
-            )
-            if image_url:
-                msg_dict["content"] = [
-                    {"type": "text", "text": msg.content},
-                    {"type": "image_url", "image_url": {"url": image_url}},
-                ]
-
         conversation.append(msg_dict)
 
     tool_iterations = 0
@@ -779,11 +814,12 @@ Always be accurate with time calculations based on the current date provided.{ca
                                     msg.get("content"), pending
                                 )
                             )
-                            executed_calls, tool_results = await _execute_pending_tools(
+                            executed_calls, tool_results, extra_msgs = await _execute_pending_tools(
                                 pending, request, allowed_cameras
                             )
                             stream_tool_calls.extend(executed_calls)
                             conversation.extend(tool_results)
+                            conversation.extend(extra_msgs)
                             yield (
                                 json.dumps(
                                     {
@@ -890,11 +926,12 @@ Always be accurate with time calculations based on the current date provided.{ca
                 f"Tool calls detected (iteration {tool_iterations}/{max_iterations}): "
                 f"{len(pending_tool_calls)} tool(s) to execute"
             )
-            executed_calls, tool_results = await _execute_pending_tools(
+            executed_calls, tool_results, extra_msgs = await _execute_pending_tools(
                 pending_tool_calls, request, allowed_cameras
             )
             tool_calls.extend(executed_calls)
             conversation.extend(tool_results)
+            conversation.extend(extra_msgs)
             logger.debug(
                 f"Added {len(tool_results)} tool result(s) to conversation. "
                 f"Continuing with next LLM call..."

frigate/api/defs/request/chat_body.py

@@ -32,13 +32,6 @@ class ChatCompletionRequest(BaseModel):
         le=10,
         description="Maximum number of tool call iterations (default: 5)",
     )
-    include_live_image: Optional[str] = Field(
-        default=None,
-        description=(
-            "If set, the current live frame from this camera is attached to the first "
-            "user message as multimodal content. Use with get_live_context for detection info."
-        ),
-    )
     stream: bool = Field(
         default=False,
         description="If true, stream the final assistant response in the body as newline-delimited JSON.",