Merge branch 'main' into vectordb_name

llama_stack/providers/utils/inference/model_registry.py

@@ -83,9 +83,37 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
     def get_llama_model(self, provider_model_id: str) -> str | None:
         return self.provider_id_to_llama_model_map.get(provider_model_id, None)
 
+    async def check_model_availability(self, model: str) -> bool:
+        """
+        Check if a specific model is available from the provider (non-static check).
+
+        This is for subclassing purposes, so providers can check if a specific
+        model is currently available for use through dynamic means (e.g., API calls).
+
+        This method should NOT check statically configured model entries in
+        `self.alias_to_provider_id_map` - that is handled separately in register_model.
+
+        Default implementation returns False (no dynamic models available).
+
+        :param model: The model identifier to check.
+        :return: True if the model is available dynamically, False otherwise.
+        """
+        return False
+
     async def register_model(self, model: Model) -> Model:
-        if not (supported_model_id := self.get_provider_model_id(model.provider_resource_id)):
-            raise UnsupportedModelError(model.provider_resource_id, self.alias_to_provider_id_map.keys())
+        # Check if model is supported in static configuration
+        supported_model_id = self.get_provider_model_id(model.provider_resource_id)
+
+        # If not found in static config, check if it's available dynamically from provider
+        if not supported_model_id:
+            if await self.check_model_availability(model.provider_resource_id):
+                supported_model_id = model.provider_resource_id
+            else:
+                # note: we cannot provide a complete list of supported models without
+                #       getting a complete list from the provider, so we return "..."
+                all_supported_models = [*self.alias_to_provider_id_map.keys(), "..."]
+                raise UnsupportedModelError(model.provider_resource_id, all_supported_models)
+
         provider_resource_id = self.get_provider_model_id(model.model_id)
         if model.model_type == ModelType.embedding:
             # embedding models are always registered by their provider model id and does not need to be mapped to a llama model
@@ -114,6 +142,7 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
                         ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR[llama_model]
                     )
 
+        # Register the model alias, ensuring it maps to the correct provider model id
         self.alias_to_provider_id_map[model.model_id] = supported_model_id
 
         return model

llama_stack/providers/utils/memory/openai_vector_store_mixin.py

@@ -5,6 +5,7 @@
 # the root directory of this source tree.
 
 import asyncio
+import json
 import logging
 import mimetypes
 import time
@@ -35,6 +36,7 @@ from llama_stack.apis.vector_io import (
     VectorStoreSearchResponse,
     VectorStoreSearchResponsePage,
 )
+from llama_stack.providers.utils.kvstore.api import KVStore
 from llama_stack.providers.utils.memory.vector_store import content_from_data_and_mime_type, make_overlapped_chunks
 
 logger = logging.getLogger(__name__)
@@ -59,26 +61,45 @@ class OpenAIVectorStoreMixin(ABC):
     # These should be provided by the implementing class
     openai_vector_stores: dict[str, dict[str, Any]]
     files_api: Files | None
+    # KV store for persisting OpenAI vector store metadata
+    kvstore: KVStore | None
 
-    @abstractmethod
     async def _save_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
         """Save vector store metadata to persistent storage."""
-        pass
+        assert self.kvstore is not None
+        key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
+        await self.kvstore.set(key=key, value=json.dumps(store_info))
+        # update in-memory cache
+        self.openai_vector_stores[store_id] = store_info
 
-    @abstractmethod
     async def _load_openai_vector_stores(self) -> dict[str, dict[str, Any]]:
         """Load all vector store metadata from persistent storage."""
-        pass
+        assert self.kvstore is not None
+        start_key = OPENAI_VECTOR_STORES_PREFIX
+        end_key = f"{OPENAI_VECTOR_STORES_PREFIX}\xff"
+        stored_data = await self.kvstore.values_in_range(start_key, end_key)
+
+        stores: dict[str, dict[str, Any]] = {}
+        for item in stored_data:
+            info = json.loads(item)
+            stores[info["id"]] = info
+        return stores
 
-    @abstractmethod
     async def _update_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
         """Update vector store metadata in persistent storage."""
-        pass
+        assert self.kvstore is not None
+        key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
+        await self.kvstore.set(key=key, value=json.dumps(store_info))
+        # update in-memory cache
+        self.openai_vector_stores[store_id] = store_info
 
-    @abstractmethod
     async def _delete_openai_vector_store_from_storage(self, store_id: str) -> None:
         """Delete vector store metadata from persistent storage."""
-        pass
+        assert self.kvstore is not None
+        key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
+        await self.kvstore.delete(key)
+        # remove from in-memory cache
+        self.openai_vector_stores.pop(store_id, None)
 
     @abstractmethod
     async def _save_openai_vector_store_file(
@@ -117,6 +138,10 @@ class OpenAIVectorStoreMixin(ABC):
         """Unregister a vector database (provider-specific implementation)."""
         pass
 
+    async def initialize_openai_vector_stores(self) -> None:
+        """Load existing OpenAI vector stores into the in-memory cache."""
+        self.openai_vector_stores = await self._load_openai_vector_stores()
+
     @abstractmethod
     async def insert_chunks(
         self,