Matching intent, not strings
Semantic search finds information by closeness of meaning rather than by literal keyword match. The premise is that different wording with the same intent should still be found.
How it differs from keyword search
Keyword search for “expense reimbursement” finds documents containing those words. A document titled “getting money back that you paid out of pocket” means the same thing but will not match.
Semantic search converts text into numeric coordinates through embeddings, and looks for documents whose coordinates are close — close position meaning close meaning. That is how it reaches the right material with no shared words at all.
Where it is used
- RAG: retrieving the material an AI answers from, the core of internal document assistants
- AI search services: results that reflect the intent of a question
- E-commerce: finding products from a loosely described requirement
- Support: surfacing similar past enquiries
Search in the generative AI era runs on this almost everywhere.
Its weakness, and the usual fix
Strong on similarity, it is poor at exact matches — part numbers, dates, amounts. Production systems therefore commonly use hybrid search, combining it with keyword search. Semantic for ambiguity, keyword for precision, both together in practice, is the summary that makes explanations of AI search and RAG click.
As a user, if an internal search or an AI service seems to understand what you meant despite different wording, this is almost certainly why.
Ordinary web search has incorporated meaning-based techniques too, so the reality is convergence rather than keyword search versus semantic search.