from __future__ import annotations import logging from enum import Enum from typing import ( TYPE_CHECKING, Any, Dict, Generator, Iterable, List, Optional, Tuple, Union, ) import numpy as np from langchain_core.documents import Document from langchain_core.vectorstores import VectorStore from langchain_community.vectorstores.utils import maximal_marginal_relevance if TYPE_CHECKING: from langchain_core.embeddings import Embeddings from pymongo.collection import Collection # Before Python 3.11 native StrEnum is not available class CosmosDBSimilarityType(str, Enum): """Cosmos DB Similarity Type as enumerator.""" COS = "COS" """CosineSimilarity""" IP = "IP" """inner - product""" L2 = "L2" """Euclidean distance""" class CosmosDBVectorSearchType(str, Enum): """Cosmos DB Vector Search Type as enumerator.""" VECTOR_IVF = "vector-ivf" """IVF vector index""" VECTOR_HNSW = "vector-hnsw" """HNSW vector index""" logger = logging.getLogger(__name__) DEFAULT_INSERT_BATCH_SIZE = 128 class AzureCosmosDBVectorSearch(VectorStore): """`Azure Cosmos DB for MongoDB vCore` vector store. To use, you should have both: - the ``pymongo`` python package installed - a connection string associated with a MongoDB VCore Cluster Example: . code-block:: python from langchain_community.vectorstores import AzureCosmosDBVectorSearch from langchain_community.embeddings.openai import OpenAIEmbeddings from pymongo import MongoClient mongo_client = MongoClient("") collection = mongo_client[""][""] embeddings = OpenAIEmbeddings() vectorstore = AzureCosmosDBVectorSearch(collection, embeddings) """ def __init__( self, collection: Collection, embedding: Embeddings, *, index_name: str = "vectorSearchIndex", text_key: str = "textContent", embedding_key: str = "vectorContent", application_name: str = "LANGCHAIN_PYTHON", ): """Constructor for AzureCosmosDBVectorSearch Args: collection: MongoDB collection to add the texts to. embedding: Text embedding model to use. index_name: Name of the Atlas Search index. text_key: MongoDB field that will contain the text for each document. embedding_key: MongoDB field that will contain the embedding for each document. """ self._collection = collection self._embedding = embedding self._index_name = index_name self._text_key = text_key self._embedding_key = embedding_key self._application_name = application_name @property def embeddings(self) -> Embeddings: return self._embedding def get_index_name(self) -> str: """Returns the index name Returns: Returns the index name """ return self._index_name @classmethod def from_connection_string( cls, connection_string: str, namespace: str, embedding: Embeddings, application_name: str = "LANGCHAIN_PYTHON", **kwargs: Any, ) -> AzureCosmosDBVectorSearch: """Creates an Instance of AzureCosmosDBVectorSearch from a Connection String Args: connection_string: The MongoDB vCore instance connection string namespace: The namespace (database.collection) embedding: The embedding utility **kwargs: Dynamic keyword arguments Returns: an instance of the vector store """ try: from pymongo import MongoClient except ImportError: raise ImportError( "Could not import pymongo, please install it with " "`pip install pymongo`." ) appname = application_name client: MongoClient = MongoClient(connection_string, appname=appname) db_name, collection_name = namespace.split(".") collection = client[db_name][collection_name] return cls(collection, embedding, **kwargs) def index_exists(self) -> bool: """Verifies if the specified index name during instance construction exists on the collection Returns: Returns True on success and False if no such index exists on the collection """ cursor = self._collection.list_indexes() index_name = self._index_name for res in cursor: current_index_name = res.pop("name") if current_index_name == index_name: return True return False def delete_index(self) -> None: """Deletes the index specified during instance construction if it exists""" if self.index_exists(): self._collection.drop_index(self._index_name) # Raises OperationFailure on an error (e.g. trying to drop # an index that does not exist) def create_index( self, num_lists: int = 100, dimensions: int = 1536, similarity: CosmosDBSimilarityType = CosmosDBSimilarityType.COS, kind: str = "vector-ivf", m: int = 16, ef_construction: int = 64, ) -> dict[str, Any]: """Creates an index using the index name specified at instance construction Setting the numLists parameter correctly is important for achieving good accuracy and performance. Since the vector store uses IVF as the indexing strategy, you should create the index only after you have loaded a large enough sample documents to ensure that the centroids for the respective buckets are faily distributed. We recommend that numLists is set to documentCount/1000 for up to 1 million documents and to sqrt(documentCount) for more than 1 million documents. As the number of items in your database grows, you should tune numLists to be larger in order to achieve good latency performance for vector search. If you're experimenting with a new scenario or creating a small demo, you can start with numLists set to 1 to perform a brute-force search across all vectors. This should provide you with the most accurate results from the vector search, however be aware that the search speed and latency will be slow. After your initial setup, you should go ahead and tune the numLists parameter using the above guidance. Args: kind: Type of vector index to create. Possible options are: - vector-ivf - vector-hnsw: available as a preview feature only, to enable visit https://learn.microsoft.com/en-us/azure/azure-resource-manager/management/preview-features num_lists: This integer is the number of clusters that the inverted file (IVF) index uses to group the vector data. We recommend that numLists is set to documentCount/1000 for up to 1 million documents and to sqrt(documentCount) for more than 1 million documents. Using a numLists value of 1 is akin to performing brute-force search, which has limited performance dimensions: Number of dimensions for vector similarity. The maximum number of supported dimensions is 2000 similarity: Similarity metric to use with the IVF index. Possible options are: - CosmosDBSimilarityType.COS (cosine distance), - CosmosDBSimilarityType.L2 (Euclidean distance), and - CosmosDBSimilarityType.IP (inner product). m: The max number of connections per layer (16 by default, minimum value is 2, maximum value is 100). Higher m is suitable for datasets with high dimensionality and/or high accuracy requirements. ef_construction: the size of the dynamic candidate list for constructing the graph (64 by default, minimum value is 4, maximum value is 1000). Higher ef_construction will result in better index quality and higher accuracy, but it will also increase the time required to build the index. ef_construction has to be at least 2 * m Returns: An object describing the created index """ # check the kind of vector search to be performed # prepare the command accordingly create_index_commands = {} if kind == CosmosDBVectorSearchType.VECTOR_IVF: create_index_commands = self._get_vector_index_ivf( kind, num_lists, similarity, dimensions ) elif kind == CosmosDBVectorSearchType.VECTOR_HNSW: create_index_commands = self._get_vector_index_hnsw( kind, m, ef_construction, similarity, dimensions ) # retrieve the database object current_database = self._collection.database # invoke the command from the database object create_index_responses: dict[str, Any] = current_database.command( create_index_commands ) return create_index_responses def _get_vector_index_ivf( self, kind: str, num_lists: int, similarity: str, dimensions: int ) -> Dict[str, Any]: command = { "createIndexes": self._collection.name, "indexes": [ { "name": self._index_name, "key": {self._embedding_key: "cosmosSearch"}, "cosmosSearchOptions": { "kind": kind, "numLists": num_lists, "similarity": similarity, "dimensions": dimensions, }, } ], } return command def _get_vector_index_hnsw( self, kind: str, m: int, ef_construction: int, similarity: str, dimensions: int ) -> Dict[str, Any]: command = { "createIndexes": self._collection.name, "indexes": [ { "name": self._index_name, "key": {self._embedding_key: "cosmosSearch"}, "cosmosSearchOptions": { "kind": kind, "m": m, "efConstruction": ef_construction, "similarity": similarity, "dimensions": dimensions, }, } ], } return command def create_filter_index( self, property_to_filter: str, index_name: str, ) -> dict[str, Any]: command = { "createIndexes": self._collection.name, "indexes": [ { "key": {property_to_filter: 1}, "name": index_name, } ], } # retrieve the database object current_database = self._collection.database # invoke the command from the database object create_index_responses: dict[str, Any] = current_database.command(command) return create_index_responses def add_texts( self, texts: Iterable[str], metadatas: Optional[List[Dict[str, Any]]] = None, **kwargs: Any, ) -> List: batch_size = kwargs.get("batch_size", DEFAULT_INSERT_BATCH_SIZE) _metadatas: Union[List, Generator] = metadatas or ({} for _ in texts) texts_batch = [] metadatas_batch = [] result_ids = [] for i, (text, metadata) in enumerate(zip(texts, _metadatas)): texts_batch.append(text) metadatas_batch.append(metadata) if (i + 1) % batch_size == 0: result_ids.extend(self._insert_texts(texts_batch, metadatas_batch)) texts_batch = [] metadatas_batch = [] if texts_batch: result_ids.extend(self._insert_texts(texts_batch, metadatas_batch)) return result_ids def _insert_texts(self, texts: List[str], metadatas: List[Dict[str, Any]]) -> List: """Used to Load Documents into the collection Args: texts: The list of documents strings to load metadatas: The list of metadata objects associated with each document Returns: """ # If the text is empty, then exit early if not texts: return [] # Embed and create the documents embeddings = self._embedding.embed_documents(texts) to_insert = [ {self._text_key: t, self._embedding_key: embedding, "metadata": m} for t, m, embedding in zip(texts, metadatas, embeddings) ] # insert the documents in Cosmos DB insert_result = self._collection.insert_many(to_insert) # type: ignore return insert_result.inserted_ids @classmethod def from_texts( cls, texts: List[str], embedding: Embeddings, metadatas: Optional[List[dict]] = None, collection: Optional[Collection] = None, **kwargs: Any, ) -> AzureCosmosDBVectorSearch: if collection is None: raise ValueError("Must provide 'collection' named parameter.") vectorstore = cls(collection, embedding, **kwargs) vectorstore.add_texts(texts, metadatas=metadatas) return vectorstore def delete(self, ids: Optional[List[str]] = None, **kwargs: Any) -> Optional[bool]: if ids is None: raise ValueError("No document ids provided to delete.") for document_id in ids: self.delete_document_by_id(document_id) return True def delete_document_by_id(self, document_id: Optional[str] = None) -> None: """Removes a Specific Document by Id Args: document_id: The document identifier """ try: from bson.objectid import ObjectId except ImportError as e: raise ImportError( "Unable to import bson, please install with `pip install bson`." ) from e if document_id is None: raise ValueError("No document id provided to delete.") self._collection.delete_one({"_id": ObjectId(document_id)}) def _similarity_search_with_score( self, embeddings: List[float], k: int = 4, kind: CosmosDBVectorSearchType = CosmosDBVectorSearchType.VECTOR_IVF, pre_filter: Optional[Dict] = None, ef_search: int = 40, score_threshold: float = 0.0, with_embedding: bool = False, ) -> List[Tuple[Document, float]]: """Returns a list of documents with their scores Args: embeddings: The query vector k: the number of documents to return kind: Type of vector index to create. Possible options are: - vector-ivf - vector-hnsw: available as a preview feature only, to enable visit https://learn.microsoft.com/en-us/azure/azure-resource-manager/management/preview-features ef_search: The size of the dynamic candidate list for search (40 by default). A higher value provides better recall at the cost of speed. score_threshold: (Optional[float], optional): Maximum vector distance between selected documents and the query vector. Defaults to None. Only vector-ivf search supports this for now. Returns: A list of documents closest to the query vector """ pipeline: List[dict[str, Any]] = [] if kind == CosmosDBVectorSearchType.VECTOR_IVF: pipeline = self._get_pipeline_vector_ivf(embeddings, k, pre_filter) elif kind == CosmosDBVectorSearchType.VECTOR_HNSW: pipeline = self._get_pipeline_vector_hnsw( embeddings, k, ef_search, pre_filter ) cursor = self._collection.aggregate(pipeline) docs = [] for res in cursor: score = res.pop("similarityScore") if score < score_threshold: continue document_object_field = res.pop("document") text = document_object_field.pop(self._text_key) metadata = document_object_field.pop("metadata", {}) if with_embedding: metadata[self._embedding_key] = document_object_field.pop( self._embedding_key ) docs.append((Document(page_content=text, metadata=metadata), score)) return docs def _get_pipeline_vector_ivf( self, embeddings: List[float], k: int = 4, pre_filter: Optional[Dict] = None ) -> List[dict[str, Any]]: params = { "vector": embeddings, "path": self._embedding_key, "k": k, } if pre_filter: params["filter"] = pre_filter pipeline: List[dict[str, Any]] = [ { "$search": { "cosmosSearch": params, "returnStoredSource": True, } }, { "$project": { "similarityScore": {"$meta": "searchScore"}, "document": "$$ROOT", } }, ] return pipeline def _get_pipeline_vector_hnsw( self, embeddings: List[float], k: int = 4, ef_search: int = 40, pre_filter: Optional[Dict] = None, ) -> List[dict[str, Any]]: params = { "vector": embeddings, "path": self._embedding_key, "k": k, "efSearch": ef_search, } if pre_filter: params["filter"] = pre_filter pipeline: List[dict[str, Any]] = [ { "$search": { "cosmosSearch": params, } }, { "$project": { "similarityScore": {"$meta": "searchScore"}, "document": "$$ROOT", } }, ] return pipeline def similarity_search_with_score( self, query: str, k: int = 4, kind: CosmosDBVectorSearchType = CosmosDBVectorSearchType.VECTOR_IVF, pre_filter: Optional[Dict] = None, ef_search: int = 40, score_threshold: float = 0.0, with_embedding: bool = False, ) -> List[Tuple[Document, float]]: embeddings = self._embedding.embed_query(query) docs = self._similarity_search_with_score( embeddings=embeddings, k=k, kind=kind, pre_filter=pre_filter, ef_search=ef_search, score_threshold=score_threshold, with_embedding=with_embedding, ) return docs def similarity_search( self, query: str, k: int = 4, kind: CosmosDBVectorSearchType = CosmosDBVectorSearchType.VECTOR_IVF, pre_filter: Optional[Dict] = None, ef_search: int = 40, score_threshold: float = 0.0, with_embedding: bool = False, **kwargs: Any, ) -> List[Document]: docs_and_scores = self.similarity_search_with_score( query, k=k, kind=kind, pre_filter=pre_filter, ef_search=ef_search, score_threshold=score_threshold, with_embedding=with_embedding, ) return [doc for doc, _ in docs_and_scores] def max_marginal_relevance_search_by_vector( self, embedding: List[float], k: int = 4, fetch_k: int = 20, lambda_mult: float = 0.5, kind: CosmosDBVectorSearchType = CosmosDBVectorSearchType.VECTOR_IVF, pre_filter: Optional[Dict] = None, ef_search: int = 40, score_threshold: float = 0.0, with_embedding: bool = False, **kwargs: Any, ) -> List[Document]: # Retrieves the docs with similarity scores # sorted by similarity scores in DESC order docs = self._similarity_search_with_score( embedding, k=fetch_k, kind=kind, pre_filter=pre_filter, ef_search=ef_search, score_threshold=score_threshold, with_embedding=with_embedding, ) # Re-ranks the docs using MMR mmr_doc_indexes = maximal_marginal_relevance( np.array(embedding), [doc.metadata[self._embedding_key] for doc, _ in docs], k=k, lambda_mult=lambda_mult, ) mmr_docs = [docs[i][0] for i in mmr_doc_indexes] return mmr_docs def max_marginal_relevance_search( self, query: str, k: int = 4, fetch_k: int = 20, lambda_mult: float = 0.5, kind: CosmosDBVectorSearchType = CosmosDBVectorSearchType.VECTOR_IVF, pre_filter: Optional[Dict] = None, ef_search: int = 40, score_threshold: float = 0.0, with_embedding: bool = False, **kwargs: Any, ) -> List[Document]: # compute the embeddings vector from the query string embeddings = self._embedding.embed_query(query) docs = self.max_marginal_relevance_search_by_vector( embeddings, k=k, fetch_k=fetch_k, lambda_mult=lambda_mult, kind=kind, pre_filter=pre_filter, ef_search=ef_search, score_threshold=score_threshold, with_embedding=with_embedding, ) return docs def get_collection(self) -> Collection: return self._collection