Semantic search
Search that ranks results by meaning rather than keyword overlap — using vector embeddings or LLM reasoning to match queries with conceptually similar content.
Semantic search is what users actually want from search but did not get for thirty years. "Documents about renewing TLS certificates" matches a page titled "SSL certificate rotation procedures" — different words, same meaning. Traditional keyword search fails; semantic search succeeds.
The 2026 production implementation: hybrid retrieval. Vector search for semantic matches, BM25 or similar for keyword precision, fused scores, and a cross-encoder reranker for the top results. Pure-vector retrieval has been beaten on every major benchmark by hybrid approaches.
For agents, semantic search is the retrieval layer of RAG. Documents get embedded once; queries get matched at search time; the agent reads the top results. Get the chunking, embedding model, and reranker right and the rest of the agent stack is downstream from that quality.
Frequently asked
What is the difference between semantic search and keyword search?+
Keyword search matches exact words. Semantic search matches meanings. "Car insurance quotes" and "auto policy estimates" mean similar things; keyword search treats them as unrelated, semantic search treats them as near-synonyms.
Is semantic search always better than keyword search?+
For natural-language queries: usually. For exact-string matching (error codes, product SKUs, proper nouns): no — keyword search beats semantic search on precision. Hybrid retrieval gets the best of both.