On this page
Definition
Vector search is a retrieval technique that ranks results by similarity between numeric embeddings, enabling meaning-based (semantic) matching rather than lexical matching.
Simple explanation
Keyword search asks 'which documents contain this exact word?'. Vector search asks 'which documents mean roughly the same as this query?'.
That means a search for 'how do I get a refund' can surface a policy titled 'Money-back returns' even though the words are different.
Why it matters
It underpins nearly every modern AI search product and is the retrieval half of RAG.
How it works
- 1EmbedConvert each document into a vector via an embedding model.
- 2StoreSave vectors in a database optimised for nearest-neighbour lookup.
- 3QueryEmbed the user's question and search for the nearest vectors.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- pgvectorAdds vector search to Postgres.
- Pinecone, Weaviate, QdrantManaged vector databases.
Advantages
- Understands intent, not just wording.
- Handles typos and paraphrasing.
- Multilingual.
Limitations
- Quality depends on the embedding model.
- Not always better than keyword search alone.
- Requires additional infrastructure.
Common misunderstandings
- ClaimVector search replaces keyword search.RealityThe best systems combine both (hybrid search).
Frequently asked questions
Do I need a special database?
For scale yes; for small datasets a normal database plus a vector extension works fine.
Is vector search deterministic?
Yes for a given index and query, though results shift when documents are re-embedded.
The Tool Money Lab perspective
Vector search is the boring infrastructure that makes AI search feel magical. It is worth learning even if you never touch the code.
Conclusion
Vector search moves retrieval from 'exact words' to 'shared meaning' — the foundation of most useful AI over private data.