MongoDB Atlas 集成

Integrate with the MongoDB Atlas Vector Store using LangChain Python.

This notebook covers how to use MongoDB Vector Search with LangChain. It also showcases the MongoDB Atlas Embedding and Reranking API for accessing Voyage AI’s state-of-the-art embedding models and rerankers.

MongoDB Atlas is a fully-managed cloud database available in AWS, Azure, and GCP. It supports native Vector Search, full text search (BM25), and hybrid search on your MongoDB document data.

MongoDB Vector Search allows to store your embeddings in MongoDB documents, create a vector search index, and perform KNN search with an approximate nearest neighbor algorithm (Hierarchical Navigable Small Worlds). It uses the $vectorSearch MQL Stage.

MongoDB Atlas Embedding and Reranking API enables access to Voyage AI models via MongoDB Atlas.

Note: The MongoDBAtlasVectorSearch interface is compatible with MongoDB Vector Search in the Community Edition.

设置

To use MongoDB Atlas, you must first deploy a cluster. To get started, sign up for free to Atlas.

In order to use Voyage AI embedding and reranking models, you will need to create a model API key. Generate your API key, and get access to 200 million free tokens on the latest models.

First, start by installing the following libraries to follow this notebook.

!pip install -qU langchain-mongodb langchain-voyageai langchain-community

Replace the placeholder values in the following code cell with your Atlas Cluster connection string and your Model API Key.

import os
from pymongo import MongoClient
from langchain_mongodb import MongoDBAtlasVectorSearch
from langchain_mongodb.retrievers import MongoDBAtlasHybridSearchRetriever
from langchain_voyageai import VoyageAIEmbeddings, VoyageAIRerank
from langchain_community.document_loaders import WikipediaLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter

# Setup credentials
MONGODB_ATLAS_URI = "YOUR_ATLAS_CONNECTION_STRING_HERE"
MONGODB_ATLAS_MODEL_API_KEY = "YOUR_ATLAS_MODEL_API_KEY" 
os.environ["VOYAGE_API_KEY"] = MONGODB_ATLAS_MODEL_API_KEY

client = MongoClient(MONGODB_ATLAS_URI)
db_name = "langchain_db"
collection_name = "vector_store_example"
atlas_collection = client[db_name][collection_name]
vector_index_name = "vector_index"

Insert data

We load documents, generate embeddings via the Atlas-hosted Voyage AI model, and programmatically create the Vector Search index.

# Load sample data
loader = WikipediaLoader(query="MongoDB", load_max_docs=2)
data = loader.load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
docs = text_splitter.split_documents(data)

# Initialize Embeddings (Voyage AI hosted on Atlas)
embeddings = VoyageAIEmbeddings(model="voyage-4-lite")

# Initialize Vector Store
vector_store = MongoDBAtlasVectorSearch(
    collection=atlas_collection,
    embedding=embeddings,
    index_name=vector_index_name
)

Create vector index

# Create the vector search index with specific filters for optimized metadata searching
vector_store.create_vector_search_index(
    dimensions=1024,
    wait_until_complete=45 # waits 45 seconds for index readiness
)

[OPTIONAL] Alternative to the vector_store.create_vector_search_index command above, you can also create the vector search index using the Atlas UI with the following index definition:

{
  "fields":[
    {
      "type": "vector",
      "path": "embedding",
      "numDimensions": 1024,
      "similarity": "cosine"
    }
  ]
}

# Ingest documents
vector_store.add_documents(docs)

查询向量存储

Semantic search

Find the most relevant documents based on semantic similarity.

query = "What is MongoDB Atlas?"
results = vector_store.similarity_search(query, k=3)

for doc in results:
    print(f"* {doc.page_content[:100]}...")

Semantic search with score

Retrieve documents along with their relevance scores.

results_with_score = vector_store.similarity_search_with_score(query, k=3)

for doc, score in results_with_score:
    print(f"Score: {score:.4f} | Content: {doc.page_content[:80]}...")

Semantic search with filtering

First, let’s update the vector search index by providing the field to filter on.

# Create the vector search index with specific filters for optimized metadata searching
vector_store.create_vector_search_index(
    dimensions=1024,
    wait_until_complete=60,
    update=True,
    filters=["source"]
)

Narrow down results using metadata filters. Note that Atlas Vector Search requires explicit operators like $eq.

# Use a source known to exist in our loaded docs
sample_source = "https://en.wikipedia.org/wiki/MongoDB"

filtered_results = vector_store.similarity_search(
    query,
    k=3,
    pre_filter={"source": {"$eq": sample_source}}
)

print(f"Retrieved {len(filtered_results)} documents matching source: {sample_source}")

Hybrid search

Combining Vector Search with Full-Text Search (Keyword) using Reciprocal Rank Fusion (RRF).

from langchain_mongodb.index import create_fulltext_search_index

# Use helper method to create the search index
create_fulltext_search_index(
   collection = atlas_collection,
    wait_until_complete=60,
   field = "text",
   index_name = "search_index"
)

# Requires a standard Search Index named 'default' on the collection
retriever = MongoDBAtlasHybridSearchRetriever(
    vectorstore=vector_store,
    search_index_name="search_index", 
    fulltext_penalty = 50,
    vector_penalty = 50,
    top_k=5
)

hybrid_docs = retriever.invoke("database-as-a-service")
for doc in hybrid_docs:
   print("Title: " + doc.metadata["title"])
   print("Plot: " + doc.page_content)
   print("Search score: {}".format(doc.metadata["fulltext_score"]))
   print("Vector Search score: {}".format(doc.metadata["vector_score"]))
   print("Total score: {}\n".format(doc.metadata["fulltext_score"] + doc.metadata["vector_score"]))

Vector search and reranking

A two-stage process: broad recall followed by high-precision reranking to ensure maximum relevance.

from langchain_classic.retrievers.contextual_compression import ContextualCompressionRetriever

# 1. Define the base retriever to fetch a broad pool
base_retriever = vector_store.as_retriever(search_kwargs={"k": 20})

# 2. Initialize the Atlas-hosted Voyage AI Reranker
reranker = VoyageAIRerank(model="rerank-2.5-lite", top_k=3)

# 3. Create the Compression Retriever
compression_retriever = ContextualCompressionRetriever(
    base_compressor=reranker, 
    base_retriever=base_retriever
)

print("--- Reranked Results ---")
final_docs = compression_retriever.invoke("Explain how MongoDB stores vectors")
for doc in final_docs:
    print(f"Relevance Score: {doc.metadata['relevance_score']:.4f} | {doc.page_content[:100]}...")

Retrieval-Augmented Generation (RAG)

For guides on how to use the MongoDB Vector Store integration with LangChain for Retrieval-Augmented Generation (RAG), see the following tutorials:

• Basic RAG
• RAG with Hybrid Search

Other notes

• More documentation can be found at MongoDB’s LangChain Docs.
• This feature is Generally Available and ready for production deployments.

---

API 参考

For detailed documentation of all MongoDBAtlasVectorSearch features and configurations head to the API reference.

---

连接这些文档到 Claude、VSCode 等工具，通过 MCP 获取实时答案。

在 GitHub 上编辑此页面或提交 issue。