Vector database
A vector database stores numeric embeddings and finds the ones closest to a query, the index behind semantic search and RAG.
Category
Industry concept
See all terms
What is a vector database?
A vector database stores embeddings, numeric representations of text, images, or other data, and searches them by similarity instead of by exact match. Each item becomes a vector, a list of numbers encoding its meaning, and vectors that sit close together represent items that are semantically related, so a search for "puppy" can return a result about a dog even if that word never appears in it.
This solves a real limit of keyword search: exact-match search only finds text sharing the same words as the query and misses anything phrased differently even when the meaning matches. A vector database matches on meaning instead of wording, which is why it became the standard index behind semantic search and RAG.
How does a vector database differ from a context layer?
A vector database only stores embeddings, not the knowledge itself. It can tell you which chunk of text sits closest to a query, but it has no concept of who is allowed to see that chunk or whether the source document behind it has changed. A context layer holds the actual knowledge, not a numeric representation of it, in one structured, versioned, permission-aware repository that every agent reads from directly.
When a source document changes, a vector database only reflects that once a separate pipeline re-embeds and re-indexes it. A context layer updates on its own: file versions, change reviews, and continuous updates keep the repository current without a person or a pipeline doing it by hand.
How does a vector database work?
A vector database compares a query vector against stored vectors and returns the closest ones by a distance measure such as cosine similarity. Comparing a query against every stored vector one by one is too slow at real scale, so most vector databases search an approximate index instead, commonly a layered graph structure, trading a small amount of accuracy for speed measured in milliseconds instead of seconds.
That speed only helps if the underlying data is worth finding. Qontext's context repository keeps the source knowledge structured, current, and access-controlled, so anything a vector index surfaces from it reflects what's actually true right now, not a copy that drifted out of sync.