vector store

A vector store, or vector database, is the library that a RAG system searches. It holds the embedding of every chunk together with the original text and some metadata, and it is built to answer one question extremely fast: which stored vectors are closest to this query vector? Doing that across millions of entries in a few milliseconds is exactly what an ordinary database is bad at and a vector store is designed for.

Under the hood it indexes vectors with structures like HNSW graphs or IVF clustering, trading a little exactness for a lot of speed — the results are approximate but arrive almost instantly. Beyond raw similarity it usually supports metadata filtering, so you can restrict a search by date, source, or user permissions before ranking. FAISS, pgvector, and various managed services are common choices. The vector store is what makes searching by meaning practical at real-world scale.