Vector Database

A vector database stores embeddings and retrieves by nearest-neighbour similarity, so a search for “affordable family car” can surface “budget SUV” even with no shared keywords. I ran this in production on Postgres with the pgvector extension, which kept the vectors next to the relational data instead of in a separate system.

It’s the retrieval engine under most RAG. The interesting trade-offs are practical — index type, recall versus latency, how fresh the vectors stay — and they decide whether the whole system feels sharp or sluggish.