A vector database stores embeddings and finds the nearest ones to a query vector in milliseconds — the retrieval engine behind RAG and semantic search.
The options
| DB | Best for |
|---|---|
| Chroma | Dead-simple, embedded, great for prototyping |
| Qdrant | Production, fast, rich filtering, runs in Docker |
| pgvector | You already run Postgres; keep it all in SQL |
| FAISS | Library, not a server; max speed, you manage storage |
Qdrant in 30 seconds
docker run -p 6333:6333 qdrant/qdrantfrom qdrant_client import QdrantClient
c = QdrantClient("localhost", port=6333)
c.create_collection("docs", vectors_config={"size":768,"distance":"Cosine"})
c.upsert("docs", points=[{"id":1,"vector":emb,"payload":{"text":chunk}}])
hits = c.search("docs", query_vector=q_emb, limit=5)Key concepts
- Distance metric — cosine for most text embeddings; match what your model was trained on.
- HNSW index — the graph algorithm that makes search sub-linear.
- Payload filtering — combine vector search with metadata (date, source) for hybrid retrieval.
Related: RAG with Local Embeddings · Container and Docker Security