A vector database stores and indexes vector embeddings — the numeric representations that encode the meaning of a piece of text — so they can be searched by similarity at speed, even across millions of passages.
Why it matters
Finding the right passage in a large document (or a large collection of documents) means comparing a query against every stored chunk, thousands or millions of times per second. A general-purpose database isn't built for that; it stores structured rows, not clouds of high-dimensional numbers. A vector database is purpose-built for it: it organises embeddings spatially, so a nearest-neighbour lookup is fast and cheap rather than an exhaustive scan.
That speed is what makes semantic search practical. When you ask a question, the query is embedded and the vector database finds the chunks whose vectors point in the same direction — those are the passages most likely to answer it. Without an efficient index, even a perfectly accurate embedding model can't return results in a reasonable time.
The passages a vector database returns are the raw material for retrieval-augmented generation: retrieve the right chunks first, then hand them to the model to compose an answer from real source text rather than memory.
Sidenote uses vector search under the hood every time you ask a question about a document. The passages surfaced by that search are what ground every answer in the text you're actually reading — and what make it possible to cite the exact sentence behind each claim.