Don't require download check

docs/docs/configuration/semantic_search.md

@@ -76,6 +76,40 @@ Switching between V1 and V2 requires reindexing your embeddings. The embeddings
 
 :::
 
+### GenAI Provider (llama.cpp)
+
+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).
+
+To use llama.cpp for semantic search:
+
+1. Configure a GenAI provider in your config with `embeddings` in its `roles`.
+2. Set `semantic_search.model` to the GenAI config key (e.g. `default`).
+3. Start the llama.cpp server with `--embeddings` and `--mmproj` for image support:
+
+```yaml
+genai:
+  default:
+    provider: llamacpp
+    base_url: http://localhost:8080
+    model: your-model-name
+    roles:
+      - embeddings
+      - vision
+      - tools
+
+semantic_search:
+  enabled: True
+  model: default
+```
+
+The llama.cpp server must be started with `--embeddings` for the embeddings API, and `--mmproj <mmproj.gguf>` when using image embeddings. 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.
+
+:::
+
 ### GPU Acceleration
 
 The CLIP models are downloaded in ONNX format, and the `large` model can be accelerated using GPU hardware, when available. This depends on the Docker build that is used. You can also target a specific device in a multi-GPU installation.

frigate/api/app.py

@@ -38,6 +38,7 @@ from frigate.config.camera.updater import (
     CameraConfigUpdateTopic,
 )
 from frigate.ffmpeg_presets import FFMPEG_HWACCEL_VAAPI, _gpu_selector
+from frigate.genai import GenAIClientManager
 from frigate.jobs.media_sync import (
     get_current_media_sync_job,
     get_media_sync_job_by_id,
@@ -432,6 +433,7 @@ def config_set(request: Request, body: AppConfigSetBody):
     if body.requires_restart == 0 or body.update_topic:
         old_config: FrigateConfig = request.app.frigate_config
         request.app.frigate_config = config
+        request.app.genai_manager = GenAIClientManager(config)
 
         if body.update_topic:
             if body.update_topic.startswith("config/cameras/"):

frigate/api/auth.py

@@ -1037,4 +1037,4 @@ async def get_allowed_cameras_for_filter(request: Request):
     role = current_user["role"]
     all_camera_names = set(request.app.frigate_config.cameras.keys())
     roles_dict = request.app.frigate_config.auth.roles
-    return User.get_allowed_cameras(role, roles_dict, all_camera_names)
+    return User.get_allowed_cameras(role, roles_dict, all_camera_names)

frigate/api/chat.py

@@ -23,7 +23,6 @@ from frigate.api.defs.response.chat_response import (
 )
 from frigate.api.defs.tags import Tags
 from frigate.api.event import events
-from frigate.genai import get_genai_client
 
 logger = logging.getLogger(__name__)
 
@@ -383,7 +382,7 @@ async def chat_completion(
     6. Repeats until final answer
     7. Returns response to user
     """
-    genai_client = get_genai_client(request.app.frigate_config)
+    genai_client = request.app.genai_manager.tool_client
     if not genai_client:
         return JSONResponse(
             content={

frigate/api/fastapi_app.py

@@ -33,6 +33,7 @@ from frigate.comms.event_metadata_updater import (
 from frigate.config import FrigateConfig
 from frigate.config.camera.updater import CameraConfigUpdatePublisher
 from frigate.embeddings import EmbeddingsContext
+from frigate.genai import GenAIClientManager
 from frigate.ptz.onvif import OnvifController
 from frigate.stats.emitter import StatsEmitter
 from frigate.storage import StorageMaintainer
@@ -134,6 +135,7 @@ def create_fastapi_app(
     app.include_router(record.router)
     # App Properties
     app.frigate_config = frigate_config
+    app.genai_manager = GenAIClientManager(frigate_config)
     app.embeddings = embeddings
     app.detected_frames_processor = detected_frames_processor
     app.storage_maintainer = storage_maintainer

frigate/api/review.py

@@ -33,7 +33,6 @@ from frigate.api.defs.response.review_response import (
     ReviewSummaryResponse,
 )
 from frigate.api.defs.tags import Tags
-from frigate.config import FrigateConfig
 from frigate.embeddings import EmbeddingsContext
 from frigate.models import Recordings, ReviewSegment, UserReviewStatus
 from frigate.review.types import SeverityEnum
@@ -747,9 +746,7 @@ async def set_not_reviewed(
     description="Use GenAI to summarize review items over a period of time.",
 )
 def generate_review_summary(request: Request, start_ts: float, end_ts: float):
-    config: FrigateConfig = request.app.frigate_config
-
-    if not config.genai.provider:
+    if not request.app.genai_manager.vision_client:
         return JSONResponse(
             content=(
                 {

frigate/config/camera/genai.py

@@ -6,7 +6,7 @@ from pydantic import Field
 from ..base import FrigateBaseModel
 from ..env import EnvString
 
-__all__ = ["GenAIConfig", "GenAIProviderEnum"]
+__all__ = ["GenAIConfig", "GenAIProviderEnum", "GenAIRoleEnum"]
 
 
 class GenAIProviderEnum(str, Enum):
@@ -17,6 +17,12 @@ class GenAIProviderEnum(str, Enum):
     llamacpp = "llamacpp"
 
 
+class GenAIRoleEnum(str, Enum):
+    tools = "tools"
+    vision = "vision"
+    embeddings = "embeddings"
+
+
 class GenAIConfig(FrigateBaseModel):
     """Primary GenAI Config to define GenAI Provider."""
 
@@ -24,6 +30,14 @@ class GenAIConfig(FrigateBaseModel):
     base_url: Optional[str] = Field(default=None, title="Provider base url.")
     model: str = Field(default="gpt-4o", title="GenAI model.")
     provider: GenAIProviderEnum | None = Field(default=None, title="GenAI provider.")
+    roles: list[GenAIRoleEnum] = Field(
+        default_factory=lambda: [
+            GenAIRoleEnum.embeddings,
+            GenAIRoleEnum.vision,
+            GenAIRoleEnum.tools,
+        ],
+        title="GenAI roles (tools, vision, embeddings); one provider per role.",
+    )
     provider_options: dict[str, Any] = Field(
         default={}, title="GenAI Provider extra options."
     )

frigate/config/classification.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Dict, List, Optional
+from typing import Dict, List, Optional, Union
 
 from pydantic import ConfigDict, Field
 
@@ -128,9 +128,10 @@ class SemanticSearchConfig(FrigateBaseModel):
     reindex: Optional[bool] = Field(
         default=False, title="Reindex all tracked objects on startup."
     )
-    model: Optional[SemanticSearchModelEnum] = Field(
+    model: Optional[Union[SemanticSearchModelEnum, str]] = Field(
         default=SemanticSearchModelEnum.jinav1,
-        title="The CLIP model to use for semantic search.",
+        title="The CLIP model or GenAI provider name for semantic search.",
+        description="Use 'jinav1', 'jinav2' for ONNX models, or a GenAI config key (e.g. 'default') when that provider has the embeddings role.",
     )
     model_size: str = Field(
         default="small", title="The size of the embeddings model used."

frigate/config/config.py

@@ -45,7 +45,7 @@ from .camera.audio import AudioConfig
 from .camera.birdseye import BirdseyeConfig
 from .camera.detect import DetectConfig
 from .camera.ffmpeg import FfmpegConfig
-from .camera.genai import GenAIConfig
+from .camera.genai import GenAIConfig, GenAIRoleEnum
 from .camera.motion import MotionConfig
 from .camera.notification import NotificationConfig
 from .camera.objects import FilterConfig, ObjectConfig
@@ -347,9 +347,9 @@ class FrigateConfig(FrigateBaseModel):
         default_factory=ModelConfig, title="Detection model configuration."
     )
 
-    # GenAI config
-    genai: GenAIConfig = Field(
-        default_factory=GenAIConfig, title="Generative AI configuration."
+    # GenAI config (named provider configs: name -> GenAIConfig)
+    genai: Dict[str, GenAIConfig] = Field(
+        default_factory=dict, title="Generative AI configuration (named providers)."
     )
 
     # Camera config
@@ -431,6 +431,34 @@ class FrigateConfig(FrigateBaseModel):
         # set notifications state
         self.notifications.enabled_in_config = self.notifications.enabled
 
+        # validate genai: each role (tools, vision, embeddings) at most once
+        role_to_name: dict[GenAIRoleEnum, str] = {}
+        for name, genai_cfg in self.genai.items():
+            for role in genai_cfg.roles:
+                if role in role_to_name:
+                    raise ValueError(
+                        f"GenAI role '{role.value}' is assigned to both "
+                        f"'{role_to_name[role]}' and '{name}'; each role must have "
+                        "exactly one provider."
+                    )
+                role_to_name[role] = name
+
+        # validate semantic_search.model when it is a GenAI provider name
+        if self.semantic_search.enabled and isinstance(
+            self.semantic_search.model, str
+        ):
+            if self.semantic_search.model not in self.genai:
+                raise ValueError(
+                    f"semantic_search.model '{self.semantic_search.model}' is not a "
+                    "valid GenAI config key. Must match a key in genai config."
+                )
+            genai_cfg = self.genai[self.semantic_search.model]
+            if GenAIRoleEnum.embeddings not in genai_cfg.roles:
+                raise ValueError(
+                    f"GenAI provider '{self.semantic_search.model}' must have "
+                    "'embeddings' in its roles for semantic search."
+                )
+
         # set default min_score for object attributes
         for attribute in self.model.all_attributes:
             if not self.objects.filters.get(attribute):

frigate/detectors/detection_runners.py

@@ -603,4 +603,4 @@ def get_optimized_runner(
             provider_options=options,
         ),
         model_type=model_type,
-    )
+    )

frigate/embeddings/embeddings.py

@@ -28,6 +28,7 @@ from frigate.types import ModelStatusTypesEnum
 from frigate.util.builtin import EventsPerSecond, InferenceSpeed, serialize
 from frigate.util.file import get_event_thumbnail_bytes
 
+from .genai_embedding import GenAIEmbedding
 from .onnx.jina_v1_embedding import JinaV1ImageEmbedding, JinaV1TextEmbedding
 from .onnx.jina_v2_embedding import JinaV2Embedding
 
@@ -73,11 +74,13 @@ class Embeddings:
         config: FrigateConfig,
         db: SqliteVecQueueDatabase,
         metrics: DataProcessorMetrics,
+        genai_manager=None,
     ) -> None:
         self.config = config
         self.db = db
         self.metrics = metrics
         self.requestor = InterProcessRequestor()
+        self.genai_manager = genai_manager
 
         self.image_inference_speed = InferenceSpeed(self.metrics.image_embeddings_speed)
         self.image_eps = EventsPerSecond()
@@ -104,7 +107,27 @@ class Embeddings:
                 },
             )
 
-        if self.config.semantic_search.model == SemanticSearchModelEnum.jinav2:
+        model_cfg = self.config.semantic_search.model
+        is_genai_model = isinstance(model_cfg, str)
+
+        if is_genai_model:
+            embeddings_client = (
+                genai_manager.embeddings_client if genai_manager else None
+            )
+            if not embeddings_client:
+                raise ValueError(
+                    f"semantic_search.model is '{model_cfg}' (GenAI provider) but "
+                    "no embeddings client is configured. Ensure the GenAI provider "
+                    "has 'embeddings' in its roles."
+                )
+            self.embedding = GenAIEmbedding(embeddings_client)
+            self.text_embedding = lambda input_data: self.embedding(
+                input_data, embedding_type="text"
+            )
+            self.vision_embedding = lambda input_data: self.embedding(
+                input_data, embedding_type="vision"
+            )
+        elif model_cfg == SemanticSearchModelEnum.jinav2:
             # Single JinaV2Embedding instance for both text and vision
             self.embedding = JinaV2Embedding(
                 model_size=self.config.semantic_search.model_size,
@@ -118,7 +141,8 @@ class Embeddings:
             self.vision_embedding = lambda input_data: self.embedding(
                 input_data, embedding_type="vision"
             )
-        else:  # Default to jinav1
+        else:
+            # Default to jinav1
             self.text_embedding = JinaV1TextEmbedding(
                 model_size=config.semantic_search.model_size,
                 requestor=self.requestor,
@@ -136,8 +160,11 @@ class Embeddings:
         self.metrics.text_embeddings_eps.value = self.text_eps.eps()
 
     def get_model_definitions(self):
-        # Version-specific models
-        if self.config.semantic_search.model == SemanticSearchModelEnum.jinav2:
+        model_cfg = self.config.semantic_search.model
+        if isinstance(model_cfg, str):
+            # GenAI provider: no ONNX models to download
+            models = []
+        elif model_cfg == SemanticSearchModelEnum.jinav2:
             models = [
                 "jinaai/jina-clip-v2-tokenizer",
                 "jinaai/jina-clip-v2-model_fp16.onnx"
@@ -224,6 +251,14 @@ class Embeddings:
 
         embeddings = self.vision_embedding(valid_thumbs)
 
+        if len(embeddings) != len(valid_ids):
+            logger.warning(
+                "Batch embed returned %d embeddings for %d thumbnails; skipping batch",
+                len(embeddings),
+                len(valid_ids),
+            )
+            return []
+
         if upsert:
             items = []
             for i in range(len(valid_ids)):
@@ -246,9 +281,15 @@ class Embeddings:
 
     def embed_description(
         self, event_id: str, description: str, upsert: bool = True
-    ) -> np.ndarray:
+    ) -> np.ndarray | None:
         start = datetime.datetime.now().timestamp()
-        embedding = self.text_embedding([description])[0]
+        embeddings = self.text_embedding([description])
+        if not embeddings:
+            logger.warning(
+                "Failed to generate description embedding for event %s", event_id
+            )
+            return None
+        embedding = embeddings[0]
 
         if upsert:
             self.db.execute_sql(
@@ -271,8 +312,32 @@ class Embeddings:
         # upsert embeddings one by one to avoid token limit
         embeddings = []
 
-        for desc in event_descriptions.values():
-            embeddings.append(self.text_embedding([desc])[0])
+        for eid, desc in event_descriptions.items():
+            result = self.text_embedding([desc])
+            if not result:
+                logger.warning(
+                    "Failed to generate description embedding for event %s", eid
+                )
+                continue
+            embeddings.append(result[0])
+
+        if not embeddings:
+            logger.warning("No description embeddings generated in batch")
+            return np.array([])
+
+        # Build ids list for only successful embeddings - we need to track which succeeded
+        ids = list(event_descriptions.keys())
+        if len(embeddings) != len(ids):
+            # Rebuild ids/embeddings for only successful ones (match by order)
+            ids = []
+            embeddings_filtered = []
+            for eid, desc in event_descriptions.items():
+                result = self.text_embedding([desc])
+                if result:
+                    ids.append(eid)
+                    embeddings_filtered.append(result[0])
+            ids = ids
+            embeddings = embeddings_filtered
 
         if upsert:
             ids = list(event_descriptions.keys())
@@ -314,7 +379,10 @@ class Embeddings:
 
         batch_size = (
             4
-            if self.config.semantic_search.model == SemanticSearchModelEnum.jinav2
+            if (
+                isinstance(self.config.semantic_search.model, str)
+                or self.config.semantic_search.model == SemanticSearchModelEnum.jinav2
+            )
             else 32
         )
         current_page = 1
@@ -601,6 +669,8 @@ class Embeddings:
         if trigger.type == "description":
             logger.debug(f"Generating embedding for trigger description {trigger_name}")
             embedding = self.embed_description(None, trigger.data, upsert=False)
+            if embedding is None:
+                return b""
             return embedding.astype(np.float32).tobytes()
 
         elif trigger.type == "thumbnail":
@@ -636,6 +706,8 @@ class Embeddings:
                 embedding = self.embed_thumbnail(
                     str(trigger.data), thumbnail, upsert=False
                 )
+                if embedding is None:
+                    return b""
                 return embedding.astype(np.float32).tobytes()
 
         else:

frigate/embeddings/genai_embedding.py

@@ -0,0 +1,85 @@
+"""GenAI-backed embeddings for semantic search."""
+
+import io
+import logging
+from typing import TYPE_CHECKING
+
+import numpy as np
+from PIL import Image
+
+if TYPE_CHECKING:
+    from frigate.genai import GenAIClient
+
+logger = logging.getLogger(__name__)
+
+EMBEDDING_DIM = 768
+
+
+class GenAIEmbedding:
+    """Embedding adapter that delegates to a GenAI provider's embed API.
+
+    Provides the same interface as JinaV2Embedding for semantic search:
+    __call__(inputs, embedding_type) -> list[np.ndarray]. Output embeddings are
+    normalized to 768 dimensions for Frigate's sqlite-vec schema.
+    """
+
+    def __init__(self, client: "GenAIClient") -> None:
+        self.client = client
+
+    def __call__(
+        self,
+        inputs: list[str] | list[bytes] | list[Image.Image],
+        embedding_type: str = "text",
+    ) -> list[np.ndarray]:
+        """Generate embeddings for text or images.
+
+        Args:
+            inputs: List of strings (text) or bytes/PIL images (vision).
+            embedding_type: "text" or "vision".
+
+        Returns:
+            List of 768-dim numpy float32 arrays.
+        """
+        if not inputs:
+            return []
+
+        if embedding_type == "text":
+            texts = [str(x) for x in inputs]
+            embeddings = self.client.embed(texts=texts)
+        elif embedding_type == "vision":
+            images: list[bytes] = []
+            for inp in inputs:
+                if isinstance(inp, bytes):
+                    images.append(inp)
+                elif isinstance(inp, Image.Image):
+                    buf = io.BytesIO()
+                    inp.convert("RGB").save(buf, format="JPEG")
+                    images.append(buf.getvalue())
+                else:
+                    logger.warning(
+                        "GenAIEmbedding: skipping unsupported vision input type %s",
+                        type(inp).__name__,
+                    )
+            if not images:
+                return []
+            embeddings = self.client.embed(images=images)
+        else:
+            raise ValueError(
+                f"Invalid embedding_type '{embedding_type}'. Must be 'text' or 'vision'."
+            )
+
+        result = []
+        for emb in embeddings:
+            arr = np.asarray(emb, dtype=np.float32).flatten()
+            if arr.size != EMBEDDING_DIM:
+                if arr.size > EMBEDDING_DIM:
+                    arr = arr[:EMBEDDING_DIM]
+                else:
+                    arr = np.pad(
+                        arr,
+                        (0, EMBEDDING_DIM - arr.size),
+                        mode="constant",
+                        constant_values=0,
+                    )
+            result.append(arr)
+        return result

frigate/embeddings/maintainer.py

@@ -59,7 +59,7 @@ from frigate.data_processing.real_time.license_plate import (
 from frigate.data_processing.types import DataProcessorMetrics, PostProcessDataEnum
 from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.events.types import EventTypeEnum, RegenerateDescriptionEnum
-from frigate.genai import get_genai_client
+from frigate.genai import GenAIClientManager
 from frigate.models import Event, Recordings, ReviewSegment, Trigger
 from frigate.util.builtin import serialize
 from frigate.util.file import get_event_thumbnail_bytes
@@ -116,8 +116,10 @@ class EmbeddingMaintainer(threading.Thread):
         models = [Event, Recordings, ReviewSegment, Trigger]
         db.bind(models)
 
+        self.genai_manager = GenAIClientManager(config)
+
         if config.semantic_search.enabled:
-            self.embeddings = Embeddings(config, db, metrics)
+            self.embeddings = Embeddings(config, db, metrics, self.genai_manager)
 
             # Check if we need to re-index events
             if config.semantic_search.reindex:
@@ -144,7 +146,6 @@ class EmbeddingMaintainer(threading.Thread):
         self.frame_manager = SharedMemoryFrameManager()
 
         self.detected_license_plates: dict[str, dict[str, Any]] = {}
-        self.genai_client = get_genai_client(config)
 
         # model runners to share between realtime and post processors
         if self.config.lpr.enabled:
@@ -203,12 +204,15 @@ class EmbeddingMaintainer(threading.Thread):
         # post processors
         self.post_processors: list[PostProcessorApi] = []
 
-        if self.genai_client is not None and any(
+        if self.genai_manager.vision_client is not None and any(
             c.review.genai.enabled_in_config for c in self.config.cameras.values()
         ):
             self.post_processors.append(
                 ReviewDescriptionProcessor(
-                    self.config, self.requestor, self.metrics, self.genai_client
+                    self.config,
+                    self.requestor,
+                    self.metrics,
+                    self.genai_manager.vision_client,
                 )
             )
 
@@ -246,7 +250,7 @@ class EmbeddingMaintainer(threading.Thread):
             )
             self.post_processors.append(semantic_trigger_processor)
 
-        if self.genai_client is not None and any(
+        if self.genai_manager.vision_client is not None and any(
             c.objects.genai.enabled_in_config for c in self.config.cameras.values()
         ):
             self.post_processors.append(
@@ -255,7 +259,7 @@ class EmbeddingMaintainer(threading.Thread):
                     self.embeddings,
                     self.requestor,
                     self.metrics,
-                    self.genai_client,
+                    self.genai_manager.vision_client,
                     semantic_trigger_processor,
                 )
             )

frigate/genai/__init__.py

@@ -7,15 +7,27 @@ import os
 import re
 from typing import Any, Optional
 
+import numpy as np
 from playhouse.shortcuts import model_to_dict
 
 from frigate.config import CameraConfig, FrigateConfig, GenAIConfig, GenAIProviderEnum
 from frigate.const import CLIPS_DIR
 from frigate.data_processing.post.types import ReviewMetadata
+from frigate.genai.manager import GenAIClientManager
 from frigate.models import Event
 
 logger = logging.getLogger(__name__)
 
+__all__ = [
+    "GenAIClient",
+    "GenAIClientManager",
+    "GenAIConfig",
+    "GenAIProviderEnum",
+    "PROVIDERS",
+    "load_providers",
+    "register_genai_provider",
+]
+
 PROVIDERS = {}
 
 
@@ -293,6 +305,25 @@ Guidelines:
         """Get the context window size for this provider in tokens."""
         return 4096
 
+    def embed(
+        self,
+        texts: list[str] | None = None,
+        images: list[bytes] | None = None,
+    ) -> list[np.ndarray]:
+        """Generate embeddings for text and/or images.
+
+        Returns list of numpy arrays (one per input). Expected dimension is 768
+        for Frigate semantic search compatibility.
+
+        Providers that support embeddings should override this method.
+        """
+        logger.warning(
+            "%s does not support embeddings. "
+            "This method should be overridden by the provider implementation.",
+            self.__class__.__name__,
+        )
+        return []
+
     def chat_with_tools(
         self,
         messages: list[dict[str, Any]],
@@ -352,19 +383,6 @@ Guidelines:
         }
 
 
-def get_genai_client(config: FrigateConfig) -> Optional[GenAIClient]:
-    """Get the GenAI client."""
-    if not config.genai.provider:
-        return None
-
-    load_providers()
-    provider = PROVIDERS.get(config.genai.provider)
-    if provider:
-        return provider(config.genai)
-
-    return None
-
-
 def load_providers():
     package_dir = os.path.dirname(__file__)
     for filename in os.listdir(package_dir):

frigate/genai/llama_cpp.py

@@ -1,11 +1,14 @@
 """llama.cpp Provider for Frigate AI."""
 
 import base64
+import io
 import json
 import logging
 from typing import Any, Optional
 
+import numpy as np
 import requests
+from PIL import Image
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
@@ -13,6 +16,20 @@ from frigate.genai import GenAIClient, register_genai_provider
 logger = logging.getLogger(__name__)
 
 
+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")
+        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."""
@@ -101,6 +118,104 @@ class LlamaCppClient(GenAIClient):
         """Get the context window size for llama.cpp."""
         return self.genai_config.provider_options.get("context_size", 4096)
 
+    def embed(
+        self,
+        texts: list[str] | None = None,
+        images: list[bytes] | None = None,
+    ) -> list[np.ndarray]:
+        """Generate embeddings via llama.cpp /embeddings endpoint.
+
+        Supports batch requests. Uses content format with prompt_string and
+        multimodal_data for images (PR #15108). Server must be started with
+        --embeddings and --mmproj for multimodal support.
+        """
+        if self.provider is None:
+            logger.warning(
+                "llama.cpp provider has not been initialized. Check your llama.cpp configuration."
+            )
+            return []
+
+        texts = texts or []
+        images = images or []
+        if not texts and not images:
+            return []
+
+        EMBEDDING_DIM = 768
+
+        content = []
+        for text in texts:
+            content.append({"prompt_string": text})
+        for img in images:
+            # llama.cpp uses STB which does not support WebP; convert to JPEG
+            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")
+            # prompt_string must contain <__media__> placeholder for image tokenization
+            content.append({
+                "prompt_string": "<__media__>\n",
+                "multimodal_data": [encoded],
+            })
+
+        try:
+            response = requests.post(
+                f"{self.provider}/embeddings",
+                json={"content": content},
+                timeout=self.timeout,
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            items = result.get("data", result) if isinstance(result, dict) else result
+            if not isinstance(items, list):
+                logger.warning("llama.cpp embeddings returned unexpected format")
+                return []
+
+            embeddings = []
+            for item in items:
+                emb = item.get("embedding") if isinstance(item, dict) else None
+                if emb is None:
+                    logger.warning("llama.cpp embeddings item missing embedding field")
+                    continue
+                arr = np.array(emb, dtype=np.float32)
+                orig_dim = arr.size
+                if orig_dim != EMBEDDING_DIM:
+                    if orig_dim > EMBEDDING_DIM:
+                        arr = arr[:EMBEDDING_DIM]
+                        logger.debug(
+                            "Truncated llama.cpp embedding from %d to %d dimensions",
+                            orig_dim,
+                            EMBEDDING_DIM,
+                        )
+                    else:
+                        arr = np.pad(
+                            arr,
+                            (0, EMBEDDING_DIM - orig_dim),
+                            mode="constant",
+                            constant_values=0,
+                        )
+                        logger.debug(
+                            "Padded llama.cpp embedding from %d to %d dimensions",
+                            orig_dim,
+                            EMBEDDING_DIM,
+                        )
+                embeddings.append(arr)
+            return embeddings
+        except requests.exceptions.Timeout:
+            logger.warning("llama.cpp embeddings request timed out")
+            return []
+        except requests.exceptions.RequestException as e:
+            error_detail = str(e)
+            if hasattr(e, "response") and e.response is not None:
+                try:
+                    error_detail = f"{str(e)} - Response: {e.response.text[:500]}"
+                except Exception:
+                    pass
+            logger.warning("llama.cpp embeddings error: %s", error_detail)
+            return []
+        except Exception as e:
+            logger.warning("Unexpected error in llama.cpp embeddings: %s", str(e))
+            return []
+
     def chat_with_tools(
         self,
         messages: list[dict[str, Any]],

frigate/genai/manager.py

@@ -0,0 +1,89 @@
+"""GenAI client manager for Frigate.
+
+Manages GenAI provider clients from Frigate config. Configuration is read only
+in _update_config(); no other code should read config.genai. Exposes clients
+by role: tool_client, vision_client, embeddings_client.
+"""
+
+import logging
+from typing import TYPE_CHECKING, Optional
+
+from frigate.config import FrigateConfig
+from frigate.config.camera.genai import GenAIRoleEnum
+
+if TYPE_CHECKING:
+    from frigate.genai import GenAIClient
+
+logger = logging.getLogger(__name__)
+
+
+class GenAIClientManager:
+    """Manages GenAI provider clients from Frigate config."""
+
+    def __init__(self, config: FrigateConfig) -> None:
+        self._config = config
+        self._tool_client: Optional[GenAIClient] = None
+        self._vision_client: Optional[GenAIClient] = None
+        self._embeddings_client: Optional[GenAIClient] = None
+        self._update_config()
+
+    def _update_config(self) -> None:
+        """Build role clients from current Frigate config.genai.
+
+        Called from __init__ and can be called again when config is reloaded.
+        Each role (tools, vision, embeddings) gets the client for the provider
+        that has that role in its roles list.
+        """
+        from frigate.genai import PROVIDERS, load_providers
+
+        self._tool_client = None
+        self._vision_client = None
+        self._embeddings_client = None
+
+        if not self._config.genai:
+            return
+
+        load_providers()
+
+        for _name, genai_cfg in self._config.genai.items():
+            if not genai_cfg.provider:
+                continue
+            provider_cls = PROVIDERS.get(genai_cfg.provider)
+            if not provider_cls:
+                logger.warning(
+                    "Unknown GenAI provider %s in config, skipping.",
+                    genai_cfg.provider,
+                )
+                continue
+            try:
+                client = provider_cls(genai_cfg)
+            except Exception as e:
+                logger.exception(
+                    "Failed to create GenAI client for provider %s: %s",
+                    genai_cfg.provider,
+                    e,
+                )
+                continue
+
+            for role in genai_cfg.roles:
+                if role == GenAIRoleEnum.tools:
+                    self._tool_client = client
+                elif role == GenAIRoleEnum.vision:
+                    self._vision_client = client
+                elif role == GenAIRoleEnum.embeddings:
+                    self._embeddings_client = client
+
+    @property
+    def tool_client(self) -> "Optional[GenAIClient]":
+        """Client configured for the tools role (e.g. chat with function calling)."""
+        return self._tool_client
+
+    @property
+    def vision_client(self) -> "Optional[GenAIClient]":
+        """Client configured for the vision role (e.g. review descriptions, object descriptions)."""
+        return self._vision_client
+
+    @property
+    def embeddings_client(self) -> "Optional[GenAIClient]":
+        """Client configured for the embeddings role."""
+        return self._embeddings_client

frigate/util/config.py

@@ -438,6 +438,13 @@ def migrate_018_0(config: dict[str, dict[str, Any]]) -> dict[str, dict[str, Any]
     """Handle migrating frigate config to 0.18-0"""
     new_config = config.copy()
 
+    # Migrate GenAI to new format
+    genai = new_config.get("genai")
+
+    if genai and genai.get("provider"):
+        genai["roles"] = ["embeddings", "vision", "tools"]
+        new_config["genai"] = {"default": genai}
+
     # Remove deprecated sync_recordings from global record config
     if new_config.get("record", {}).get("sync_recordings") is not None:
         del new_config["record"]["sync_recordings"]

web/src/lib/const.ts

@@ -1,3 +1,6 @@
+/** ONNX embedding models that require local model downloads. GenAI providers are not in this list. */
+export const JINA_EMBEDDING_MODELS = ["jinav1", "jinav2"] as const;
+
 export const supportedLanguageKeys = [
   "en",
   "es",

web/src/pages/Explore.tsx

@@ -23,6 +23,7 @@ import { toast } from "sonner";
 import useSWR from "swr";
 import useSWRInfinite from "swr/infinite";
 import { useDocDomain } from "@/hooks/use-doc-domain";
+import { JINA_EMBEDDING_MODELS } from "@/lib/const";
 
 const API_LIMIT = 25;
 
@@ -293,7 +294,12 @@ export default function Explore() {
   const modelVersion = config?.semantic_search.model || "jinav1";
   const modelSize = config?.semantic_search.model_size || "small";
 
-  // Text model state
+  // GenAI providers have no local models to download
+  const isGenaiEmbeddings =
+    typeof modelVersion === "string" &&
+    !(JINA_EMBEDDING_MODELS as readonly string[]).includes(modelVersion);
+
+  // Text model state (skipped for GenAI - no local models)
   const { payload: textModelState } = useModelState(
     modelVersion === "jinav1"
       ? "jinaai/jina-clip-v1-text_model_fp16.onnx"
@@ -328,6 +334,10 @@ export default function Explore() {
   );
 
   const allModelsLoaded = useMemo(() => {
+    if (isGenaiEmbeddings) {
+      return true;
+    }
+
     return (
       textModelState === "downloaded" &&
       textTokenizerState === "downloaded" &&
@@ -335,6 +345,7 @@ export default function Explore() {
       visionFeatureExtractorState === "downloaded"
     );
   }, [
+    isGenaiEmbeddings,
     textModelState,
     textTokenizerState,
     visionModelState,
@@ -358,10 +369,11 @@ export default function Explore() {
     !defaultViewLoaded ||
     (config?.semantic_search.enabled &&
       (!reindexState ||
-        !textModelState ||
-        !textTokenizerState ||
-        !visionModelState ||
-        !visionFeatureExtractorState))
+        (!isGenaiEmbeddings &&
+          (!textModelState ||
+            !textTokenizerState ||
+            !visionModelState ||
+            !visionFeatureExtractorState))))
   ) {
     return (
       <ActivityIndicator className="absolute left-1/2 top-1/2 -translate-x-1/2 -translate-y-1/2" />