Vector search
Also: Approximate nearest neighbour search · ANN
Finding the k chunks whose embeddings are closest to a query embedding. The mechanical step behind semantic search and neighbour lookup.
Last updated
Nearest-neighbour lookup in a vector index. Given a query embedding, return the k closest chunks by cosine similarity or L2 distance. Modern libraries use approximate algorithms (HNSW, IVF) that trade a little accuracy for a lot of speed, so million-document corpora feel instant.
Why it matters for your research. This is the mechanical step behind both semantic search and “show me similar passages”. When the UI says “nearest neighbours”, vector search is what happened under the hood.
In Archēglyph. Our zvec index provides the ANN primitive; results always carry the embedding model id so the researcher can tell which geometry produced the neighbours.
Not to be confused with. BM25 is not vector search — it searches an inverted word index, not an embedding space.