Vector Databases

Store embeddings and search for similar content.

Vector databases store content as embeddings and enable similarity search. When an agent searches the knowledge base, the query is converted to an embedding and matched against stored vectors to find relevant content.

How It Works

Chunk

Documents are split into smaller pieces for more precise retrieval.

Embed

Each chunk is converted to a vector embedding and stored in the database.

Search

Queries are embedded and matched against stored vectors to find similar content.

Many vector databases support a hybrid search mode that combines vector similarity with a provider-specific keyword or lexical ranking.

Hybrid search works by:

  1. Finding semantically similar content via vector search
  2. Finding lexical matches with the database's keyword-search implementation
  3. Combining lexical and vector signals using the selected database’s scoring or rank-fusion algorithm

Supported Databases

Azure Cosmos DB

MongoDB vCore vector search

Cassandra

Distributed database with vector search

Chroma

Open-source embedding database

ClickHouse

Analytical database with vector search

Couchbase

NoSQL with vector search

LanceDB

Local, serverless, hybrid search

LangChain

Use any LangChain vector store

LightRAG

Graph-based RAG

LlamaIndex

Search existing LlamaIndex indexes

Milvus

Scalable vector database

MongoDB

Atlas vector search

PgVector

PostgreSQL extension, hybrid search

Pinecone

Managed vector database

Qdrant

High-performance vector search

Redis

In-memory with vector search

Valkey

In-memory with vector search

SingleStore

Real-time analytics with vectors

SurrealDB

Multi-model database

Upstash

Serverless vector search

Weaviate

Vector search with modules

Choosing a Database

Local Development

LanceDB or ChromaDB for zero-setup local development

Production

PgVector if you already use PostgreSQL

Managed Service

Pinecone or Weaviate Cloud for vendor-operated infrastructure

Distributed Deployment

Qdrant or Milvus for a separately operated vector service

Async Support

Vector databases with async support let an async application await ingestion and search I/O. Use ainsert and asearch inside an existing event loop.

# Async insert
await knowledge.ainsert(url="https://example.com/docs.pdf")

# Async search
results = await knowledge.asearch(query="How do I configure X?")