What is vector search?
Vector search is a way of finding items by similarity of meaning. A query is converted into an embedding vector, and the system returns the stored items whose vectors are nearest to it, ranked by distance. Because the comparison happens on meaning rather than on words, a search for “engine won’t turn over” retrieves a manual section titled “starter motor troubleshooting”. Vector search is the core operation of a vector database and the retrieval step inside retrieval-augmented generation.
At scale the search uses approximate nearest neighbor algorithms, which organize vectors into graph or cluster structures so that a query inspects a small fraction of the collection. The approximation occasionally misses the true nearest item, and index settings control how often.
Vector search in practice
An engineering consultancy indexes thirty years of project reports. A structural engineer searching for past work on cold-formed steel connections in seismic zones finds relevant reports even when they used different terminology, which keyword search had never surfaced. The firm pairs the vector results with a keyword pass in a hybrid search so that project numbers and client names still match literally.
What to watch for
- Vector search alone can miss exact identifiers such as part numbers and case numbers, which is the reason hybrid search exists.
- Measure recall on a set of known question and passage pairs before tuning for speed.
- Use the same embedding model for documents and queries, or the distances will be meaningless.