Glossary
What is vector search?
In short
Vector search is search by meaning: the query and the documents are turned into sets of numbers, and the system finds the texts whose numbers are closest to the query.
Search and RAG · vector search
MMOOZZGG.com editorsChecked: 1 min read
For a child
You ask a friend: “What can we play when it rains?” Your friend does not look for the word “rain” in a book of games. They think of board games, building bricks and hide-and-seek at home, because they understood what you meant. Vector search teaches a computer to understand the meaning of a question in the same way, not just to look for the same words.
In plain words
Ordinary search looks for the words of the query in the text. If a document says “contract termination” and you ask “how to break a contract”, it may find nothing. Vector search works differently. A special model turns each fragment of text into a set of numbers that describes its meaning. The query becomes the same kind of set. Then the system takes the fragments whose numbers are closest to the numbers of the query. Phrases with similar meaning get close numbers, even if they share no words.
For business
Semantic search solves what annoys staff and customers in ordinary search: synonyms, typos and questions in their own words. Companies add it to search in a knowledge base, a catalog and support requests, and it is the first step of any assistant that answers from documents. Costs come from calculating embeddings for all documents, storing them and recalculating them when you change the model. The weak point: it finds product codes, numbers and rare terms less well, so working systems add keyword search to it. Test the quality on your real queries: how often the right document lands in the top five. An AI assistant for documents is built on this kind of search.
For a specialist
Vector search, or dense retrieval, matches the query and the documents in a shared embedding space. In the two-encoder setup (bi-encoder, dual encoder), the query and the passage are encoded separately, so the corpus embeddings are computed in advance, and relevance is scored by dot product or cosine. The 2020 paper by Karpukhin et al. on Dense Passage Retrieval showed that such a retriever, trained on a small number of question and passage pairs, beats BM25 by 9–19 points in top-20 accuracy on open-domain questions. The authors built the passage index in the FAISS library. For large corpora, search is approximate (ANN, for example HNSW). The method’s limits are weak matching on rare tokens and exact identifiers and a dependence on the domain the encoder was trained on. Hybrid search and reranking make up for them.
Example
An employee searches the knowledge base for “what to do if a customer asks for their money back”. Vector search finds the policy “Refund procedure”, even though no word of the query appears in its title.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map