Overview

A vector stores embedded data and performs similarity search.

Interface

LangChain provides a unified interface for vector stores, allowing you to:

• add_documents - Add documents to the store.
• delete - Remove stored documents by ID.
• similarity_search - Query for semantically similar documents.

This abstraction lets you switch between different implementations without altering your application logic.

Initialization

To initialize a vector store, provide it with an embedding model:

Adding documents

Add Document objects (holding page_content and optional metadata) like so:

Deleting documents

Delete by specifying IDs:

Similarity search

Issue a semantic query using similarity_search, which returns the closest embedded documents:

Many vector stores support parameters like:

• k — number of results to return
• filter — conditional filtering based on metadata

Similarity metrics & indexing

Embedding similarity may be computed using:

• Cosine similarity
• Euclidean distance
• Dot product

Efficient search often employs indexing methods such as HNSW (Hierarchical Navigable Small World), though specifics depend on the vector store.

Metadata filtering

Filtering by metadata (e.g., source, date) can refine search results:

Select embedding model:

Select vector store:

All vector stores

---