A vector database stores and searches high-dimensional vectors — the numerical embeddings that encode meaning. Where a traditional database answers "give me the row where id = 4827," a vector database answers "give me the 10 items most similar to this query." That shift in what a database does is the foundation of modern machine intelligence retrieval.

The core operation is approximate nearest neighbor (ANN) search. Documents, product descriptions, support tickets — all become embeddings at index time. At query time, the database compares your query vector against millions of stored vectors and returns the closest matches, typically in single-digit milliseconds. This is the retrieval layer that powers RAG.

The vendor landscape is noisy but the decision is usually simpler than the marketing suggests. Pinecone, Weaviate, Qdrant, and Chroma are purpose-built options. For most teams, the pragmatic starting point is pgvector — a Postgres extension that handles millions of vectors on infrastructure you already run. Purpose-built databases earn their keep at scale: hundreds of millions of vectors, sub-10ms latency requirements, complex hybrid filtering. For an MVP or a corpus under a few million records, Postgres is almost always enough.

The real decision isn't which vector database to pick. It's whether your embeddings and chunking strategy are good enough that retrieval returns the right context. The database is fast plumbing. Bad inputs just mean you're finding the wrong documents very quickly.