Vector search retrieves results by finding the stored embeddings nearest to the query's embedding, ranking by geometric distance rather than word overlap.
Also known as: nearest-neighbour search, ANN search
Exact nearest-neighbour search is too slow at scale, so production systems use approximate algorithms — HNSW graphs, IVF partitions, product quantisation — that trade a small amount of recall for very large speed gains.
Its defining strength is that it returns relevant results for queries sharing no vocabulary with the content. Its defining weakness is that it always returns something: with no keyword anchor, an unmatched query yields the nearest items in the space, which can be confidently wrong.
This is why vector search is usually deployed as one retriever among several rather than as a replacement for keyword search.