diff --git a/litellm/llms/mongodb/vector_stores/transformation.py b/litellm/llms/mongodb/vector_stores/transformation.py index 571061d39a2..2e69e35edcf 100644 --- a/litellm/llms/mongodb/vector_stores/transformation.py +++ b/litellm/llms/mongodb/vector_stores/transformation.py @@ -11,15 +11,18 @@ where the id names the index; the database and collection it covers come from litellm_params. """ -from collections.abc import Awaitable, Callable, Mapping, Sequence +from collections.abc import Callable, Mapping, Sequence from types import MappingProxyType from typing import TYPE_CHECKING, Final, NoReturn import httpx from pydantic import BaseModel, ConfigDict -import litellm -from litellm.llms.base_llm.vector_store.transformation import BaseDirectVectorStoreConfig +from litellm.llms.base_llm.vector_store.transformation import ( + BaseDirectVectorStoreConfig, + LiteLLMVectorStoreEmbeddingExecutor, + VectorStoreEmbeddingExecutor, +) from litellm.llms.mongodb.common_utils import ( DEFAULT_CONNECT_TIMEOUT_MS, DEFAULT_SERVER_SELECTION_TIMEOUT_MS, @@ -139,17 +142,13 @@ _KNOWN_MONGODB_PARAMS: Final = frozenset( class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig): def __init__( self, - embedding_fn: Callable[..., EmbeddingResponse] | None = None, - aembedding_fn: Callable[..., Awaitable[EmbeddingResponse]] | None = None, + embedding_executor: VectorStoreEmbeddingExecutor | None = None, sync_client_factory: Callable[[MongoClientKey], object] | None = None, async_client_factory: Callable[[MongoClientKey], object] | None = None, ) -> None: super().__init__() - self.embedding_fn: Final[Callable[..., EmbeddingResponse]] = ( - embedding_fn if embedding_fn is not None else litellm.embedding - ) - self.aembedding_fn: Final[Callable[..., Awaitable[EmbeddingResponse]]] = ( - aembedding_fn if aembedding_fn is not None else litellm.aembedding + self.embedding_executor: Final[VectorStoreEmbeddingExecutor] = ( + embedding_executor if embedding_executor is not None else LiteLLMVectorStoreEmbeddingExecutor() ) self.sync_client_factory: Final[Callable[[MongoClientKey], object]] = ( sync_client_factory if sync_client_factory is not None else get_sync_client @@ -350,6 +349,7 @@ class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig): vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams, litellm_logging_obj: "LiteLLMLoggingObj", litellm_params: Mapping[str, object], + embedding_executor: VectorStoreEmbeddingExecutor | None = None, timeout: float | httpx.Timeout | None = None, ) -> VectorStoreSearchResponse: self._reject_unknown_params(litellm_params) @@ -359,10 +359,10 @@ class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig): database: Final = params.require_database() collection: Final = params.require_collection() - embedding_response: Final = self.embedding_fn( - model=params.require_embedding_model(), - input=[query_text], # mutable-ok: litellm.embedding's input contract is a list - **(params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG), + embedding_response: Final = (embedding_executor or self.embedding_executor).embed( + params.require_embedding_model(), + query_text, + params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG, ) pipeline: Final = self._pipeline( vector_store_id, self._embedding_vector(embedding_response), params, vector_store_search_optional_params @@ -392,6 +392,7 @@ class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig): vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams, litellm_logging_obj: "LiteLLMLoggingObj", litellm_params: Mapping[str, object], + embedding_executor: VectorStoreEmbeddingExecutor | None = None, timeout: float | httpx.Timeout | None = None, ) -> VectorStoreSearchResponse: self._reject_unknown_params(litellm_params) @@ -401,10 +402,10 @@ class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig): database: Final = params.require_database() collection: Final = params.require_collection() - embedding_response: Final = await self.aembedding_fn( - model=params.require_embedding_model(), - input=[query_text], # mutable-ok: litellm.embedding's input contract is a list - **(params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG), + embedding_response: Final = await (embedding_executor or self.embedding_executor).aembed( + params.require_embedding_model(), + query_text, + params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG, ) pipeline: Final = self._pipeline( vector_store_id, self._embedding_vector(embedding_response), params, vector_store_search_optional_params diff --git a/tests/test_litellm/llms/mongodb/vector_stores/test_mongodb_transformation.py b/tests/test_litellm/llms/mongodb/vector_stores/test_mongodb_transformation.py index 7d71ff5c213..7a2df28cc04 100644 --- a/tests/test_litellm/llms/mongodb/vector_stores/test_mongodb_transformation.py +++ b/tests/test_litellm/llms/mongodb/vector_stores/test_mongodb_transformation.py @@ -111,27 +111,27 @@ class FakeClient: return self.database -class FakeEmbeddingFn: +class FakeEmbeddingExecutor: def __init__(self, embedding): self.embedding = embedding - self.captured_kwargs = None + self.captured = None - def __call__(self, **kwargs): - self.captured_kwargs = kwargs + def _respond(self, model, query, configuration): + self.captured = SimpleNamespace(model=model, query=query, configuration=configuration) return SimpleNamespace(data=[{"embedding": self.embedding}] if self.embedding is not None else []) + def embed(self, model, query, configuration): + return self._respond(model, query, configuration) -class FakeAsyncEmbeddingFn(FakeEmbeddingFn): - async def __call__(self, **kwargs): - self.captured_kwargs = kwargs - return SimpleNamespace(data=[{"embedding": self.embedding}] if self.embedding is not None else []) + async def aembed(self, model, query, configuration): + return self._respond(model, query, configuration) def _config(documents=(), embedding=(0.1, 0.2, 0.3), error=None, search_indexes=None): collection = FakeCollection(list(documents), error, search_indexes) client = FakeClient(collection) config = MongoDBVectorStoreConfig( - embedding_fn=FakeEmbeddingFn(list(embedding) if embedding is not None else None), + embedding_executor=FakeEmbeddingExecutor(list(embedding) if embedding is not None else None), sync_client_factory=lambda key: client, ) return config, client, collection @@ -141,7 +141,7 @@ def _async_config(documents=(), embedding=(0.1, 0.2, 0.3), error=None, search_in collection = FakeAsyncCollection(list(documents), error, search_indexes) client = FakeClient(collection) config = MongoDBVectorStoreConfig( - aembedding_fn=FakeAsyncEmbeddingFn(list(embedding) if embedding is not None else None), + embedding_executor=FakeEmbeddingExecutor(list(embedding) if embedding is not None else None), async_client_factory=lambda key: client, ) return config, client, collection @@ -252,22 +252,20 @@ def test_num_candidates_below_the_limit_or_above_the_ceiling_is_rejected(configu def test_list_query_is_joined_into_one_embedding_input(): config, _, _ = _config() - embedding_fn = config.embedding_fn _search(config, query=["deep", "space", "rescue"]) - assert embedding_fn.captured_kwargs["input"] == ["deep space rescue"] + assert config.embedding_executor.captured.query == "deep space rescue" def test_embedding_config_is_expanded_into_the_embedding_call(): config, _, _ = _config() - embedding_fn = config.embedding_fn _search(config, litellm_params={"litellm_embedding_config": {"api_base": "https://example.test", "timeout": 7}}) - assert embedding_fn.captured_kwargs["api_base"] == "https://example.test" - assert embedding_fn.captured_kwargs["timeout"] == 7 - assert embedding_fn.captured_kwargs["model"] == "openai/text-embedding-ada-002" + captured = config.embedding_executor.captured + assert captured.configuration == {"api_base": "https://example.test", "timeout": 7} + assert captured.model == "openai/text-embedding-ada-002" def test_response_maps_documents_to_openai_shaped_results(): @@ -530,7 +528,7 @@ def test_search_fails_when_the_embedding_model_returns_nothing(): def test_validation_runs_before_any_connection_is_opened(): opened = [] config = MongoDBVectorStoreConfig( - embedding_fn=FakeEmbeddingFn([0.1]), + embedding_executor=FakeEmbeddingExecutor([0.1]), sync_client_factory=lambda key: opened.append(key) or FakeClient(FakeCollection([])), ) @@ -973,7 +971,7 @@ class TestEmptyResultsAreDisambiguated: collection = ExplodingCollection([], None, []) config = MongoDBVectorStoreConfig( - embedding_fn=FakeEmbeddingFn([0.1]), + embedding_executor=FakeEmbeddingExecutor([0.1]), sync_client_factory=lambda key: FakeClient(collection), ) @@ -1157,7 +1155,7 @@ class TestClientConstructionFailures: raise error return MongoDBVectorStoreConfig( - embedding_fn=FakeEmbeddingFn([0.1, 0.2, 0.3]), sync_client_factory=factory + embedding_executor=FakeEmbeddingExecutor([0.1, 0.2, 0.3]), sync_client_factory=factory ) def _async_config_that_fails_to_connect(self, error): @@ -1165,7 +1163,7 @@ class TestClientConstructionFailures: raise error return MongoDBVectorStoreConfig( - aembedding_fn=FakeAsyncEmbeddingFn([0.1, 0.2, 0.3]), async_client_factory=factory + embedding_executor=FakeEmbeddingExecutor([0.1, 0.2, 0.3]), async_client_factory=factory ) def test_a_malformed_uri_is_a_bad_request_not_a_500(self): @@ -1215,7 +1213,7 @@ class TestSelfManagedDeploymentsAreFirstClass: raise error return MongoDBVectorStoreConfig( - embedding_fn=FakeEmbeddingFn([0.1, 0.2, 0.3]), sync_client_factory=factory + embedding_executor=FakeEmbeddingExecutor([0.1, 0.2, 0.3]), sync_client_factory=factory ) def test_a_plain_mongodb_uri_without_srv_or_credentials_is_accepted(self): @@ -1405,3 +1403,41 @@ class TestUnreadableTlsFilesAreDiagnosed: translated = translate_mongo_error(OSError("socket hung up"), index_name=INDEX, database="db", collection="c") assert not isinstance(translated, BadRequestError) + + +class TestTheCallerSuppliedEmbeddingExecutorIsUsed: + """litellm.vector_stores.search always hands a direct provider an embedding_executor, so the + provider has to accept it and route the query through it rather than its own default.""" + + def test_the_supplied_executor_produces_the_query_vector(self): + config, _, collection = _config(embedding=(0.9, 0.9, 0.9), search_indexes=READY_INDEX) + caller = FakeEmbeddingExecutor([0.4, 0.5, 0.6]) + + config.execute_search_vector_store_request( + vector_store_id=INDEX, + query="a lone astronaut", + vector_store_search_optional_params={}, + litellm_logging_obj=MagicMock(), + litellm_params=BASE_PARAMS, + embedding_executor=caller, + ) + + assert caller.captured.query == "a lone astronaut" + assert _stage(collection, "$vectorSearch")["queryVector"] == (0.4, 0.5, 0.6) + + @pytest.mark.asyncio + async def test_the_supplied_executor_produces_the_query_vector_on_the_async_path(self): + config, _, collection = _async_config(embedding=(0.9, 0.9, 0.9), search_indexes=READY_INDEX) + caller = FakeEmbeddingExecutor([0.4, 0.5, 0.6]) + + await config.aexecute_search_vector_store_request( + vector_store_id=INDEX, + query="a lone astronaut", + vector_store_search_optional_params={}, + litellm_logging_obj=MagicMock(), + litellm_params=BASE_PARAMS, + embedding_executor=caller, + ) + + assert caller.captured.query == "a lone astronaut" + assert _stage(collection, "$vectorSearch")["queryVector"] == (0.4, 0.5, 0.6)