Vector search
Vector search is finding items by the similarity of their embeddings rather than by matching words. Each document, image or passage is stored as a vector of numbers produced by a neural model, the query is turned into a vector the same way, and the engine returns the items whose vectors are closest, usually by cosine similarity or dot product. At scale, it uses approximate nearest-neighbor indexes.
It is the usual retrieval layer of semantic search and of many retrieval-augmented generation systems. Approximate nearest-neighbor methods such as HNSW (Malkov and Yashunin, 2016) and libraries such as FAISS (Johnson et al., 2017) made vector search fast enough for billions of vectors.
Example
A help center stores 50,000 articles as 768-dimensional vectors. With vector search, the question “can I change my delivery address” becomes a vector too, and the nearest article is “Edit an order after checkout”, which shares no keywords with the question.
FAQ
What is approximate nearest-neighbor search?
Approximate nearest-neighbor (ANN) search is the technique that makes vector search fast at scale: it returns vectors that are very likely, but not guaranteed, to be the closest, in a fraction of the time an exact comparison with every vector would take.
Does vector search work for exact names and IDs?
Vector search is weak at exact names, codes and IDs, because similar-looking strings can have unrelated embeddings. Systems that need both usually combine vector search with keyword search in hybrid search.
Related terms
Written by Keenable.
Published
