Semantic Search
Search that matches meaning rather than exact words, typically via embeddings or graphs.
Semantic search is a retrieval paradigm that aims to match the intent and meaning of a query rather than its literal words, so that a search for car returns results about vehicles and a search for how to fix a leaky faucet returns plumbing guidance even when the exact phrase leaky faucet does not appear. The term predates modern machine learning and originally covered knowledge-graph and ontology-based approaches, but in current usage it almost always denotes vector search driven by learned embeddings.
Mechanically, a semantic search system embeds both the query and the documents in the corpus into a shared vector space — usually with a transformer or similar neural model trained on contrastive objectives — and retrieves documents whose embeddings are close to the query's. The result is then often re-ranked, filtered, or composed with lexical signals. Some systems also use explicit semantic representations: knowledge graphs that encode entities and relations, or symbolic parsers that build structured queries from natural language. In practice these symbolic approaches are now layered on top of dense retrieval rather than used in isolation.
The advantage of semantic search is robustness to vocabulary mismatch: the user does not need to know the exact terms in the corpus. The cost is opacity and the encoding model's failure modes — the same biases, blind spots, and surprise behaviors that the underlying model has. Pure semantic search is also weaker on exact-match needs, which is why most production deployments pair it with keyword search. Performance varies sharply by domain: a general-purpose embedding model may perform poorly on a specialized medical or legal corpus without fine-tuning.
Open questions include how to make semantic search systems more controllable — the user cannot easily steer the result by adjusting terms the way they can in lexical search — and how to evaluate semantic quality at scale, since traditional relevance-judgment datasets emphasize exact-term overlap. There is also continued debate about whether semantic search will absorb keyword search, complement it, or split into domain-specific verticals with custom encoders for each.