Vector search is a retrieval method that stores content as embeddings and answers a query by finding the stored vectors nearest to the query's own embedding. It is how many AI products find relevant passages for retrieval-augmented generation. It is good at matching meaning and paraphrase, and weaker at exact strings like product codes, error messages or rare names, which is why it is often combined with keyword search.
How a vector search system works
Content is split into chunks, each chunk is embedded, and the vectors are stored in an index built for fast similarity lookups. At query time the question is embedded with the same model and the index returns the nearest neighbors. Exact comparison against every stored vector is slow at scale, so most systems use approximate nearest neighbor algorithms that trade a little accuracy for a lot of speed.
- Chunk: split documents into passages
- Embed: turn each passage into a vector
- Index: store vectors in a structure built for similarity lookups
- Query: embed the question and return the closest passages
Strengths and blind spots
Vector search handles vocabulary mismatch well. A query about 'getting paid faster' can find a passage about 'reducing days sales outstanding'. It struggles when the exact token matters: a specific API endpoint, a SKU, a version number or a new brand name the embedding model has rarely seen. Keyword retrieval handles those cases better, which is the reasoning behind hybrid search.
What this suggests for content
Write passages that are about one thing and that use both the plain-language description and the exact name of the thing. 'Tallowmere's audit log, the record of every change an admin makes' gives a vector system the concept and gives a keyword system the name.
Example
A docs search that missed error codes
Tallowmere, a fictional identity platform, moves its docs search to pure vector search. Conceptual questions get better results, but developers pasting an exact error code start getting general troubleshooting pages instead of the specific fix. Adding a keyword index alongside the vector index and merging the results restores exact-match lookups.
Common mistakes
Assuming vector search replaced keywords everywhere
Most production search systems combine methods. Exact terms still matter for names, codes and specific phrases.
Ignoring chunk boundaries
If a chunk splits an answer in half, neither half embeds well. Clear headings and self-contained sections reduce this risk.
Questions
A database built to store embeddings and run fast similarity searches over them. Many general databases have also added vector indexes.
Neither publishes its full retrieval design. Using vector similarity somewhere in the pipeline is common practice, but treat specifics about any one product as unconfirmed.
