Retrieval-Augmented Generation (RAG) is only as good as its retrieval layer. Picking the right vector database is the most important decision you'll make.
What to look for
- Recall vs latency — HNSW indexes trade memory for speed
- Hybrid search — combine dense vectors with BM25 for better results
- Filtering — metadata filters can make or break production use
Popular choices
- pgvector — great if you already run Postgres
- Qdrant — purpose-built, excellent filtering
- Pinecone — fully managed, scales well
Start simple. pgvector handles millions of vectors comfortably and gives you SQL joins for free.