Semantic Search
Semantic search retrieves results by meaning rather than keyword overlap — embedding queries and documents in one vector space and matching by similarity.
Semantic search embeds your corpus offline and the query at runtime, then returns nearest neighbors in that shared vector space, so users never have to guess a document's vocabulary. The flip side is exact tokens like error codes and SKUs, which is why mature systems add keyword search and a reranker.
Semantic search retrieves documents by meaning instead of word overlap: queries and documents are mapped into the same embedding space, and relevance becomes vector similarity.
The mechanism is simple once embeddings exist — embed the corpus offline into a vector database, embed the query at runtime, return the nearest neighbors. The payoff is robustness to phrasing: users don't need to guess the document's vocabulary. The cost is the flip side: semantic search can miss exact tokens — error codes, function names, SKUs — that old-fashioned keyword search nails, and it inherits whatever blind spots the embedding model has in your domain.
That's why mature retrieval is rarely semantic-only. Hybrid search pairs BM25 keyword retrieval with vector search, and a reranker re-sorts the merged candidates — recall from breadth, precision from the reranker. The full pattern, with when each piece earns its place, is in Hybrid Search & Reranking.
Frequently asked questions
- How is semantic search different from keyword search?
- Keyword (lexical) search matches the words themselves — great for exact identifiers, brittle for paraphrases. Semantic search matches meaning via embeddings, so 'laptop won't turn on' finds 'computer fails to boot.' The trade flips for exact strings: error codes and product SKUs are where keyword search still wins.
- Why do production systems combine both?
- Because their failure modes are complementary. Hybrid search runs lexical (BM25) and semantic retrieval together and merges the results, catching both the exact-match cases embeddings fuzz over and the paraphrases keywords miss — usually followed by a reranker to sort the merged pool precisely.
Filed under
search · embeddings · retrieval · rag
Related
- RerankingReranking is a second-pass scoring step: a cross-encoder model re-orders the top results from fast retrieval so the truly relevant few rise to the top.
- Vector DatabaseA vector database stores embeddings and answers nearest-neighbor queries fast — the retrieval layer under RAG and semantic search, using ANN indexes like HNSW.
- EmbeddingAn embedding is a vector of numbers representing text's meaning, placed so similar texts land close together — the foundation of semantic search and RAG.
- Hybrid Search & Reranking: From Top-50 Recall to Top-5 PrecisionHow production RAG combines dense and sparse search, fuses with RRF, and reranks — turning a wide candidate set into the few passages that actually answer.
- Exa vs Tavily: Web Search APIs for AI Agents (2026)Exa vs Tavily compared — neural semantic discovery vs agent-optimized RAG answers, pricing, MCP support, and which web search API fits your stack.
- RAG (Retrieval-Augmented Generation)RAG retrieves relevant documents from your own data and injects them into an LLM's prompt at query time, grounding answers in facts the model wasn't trained on.
- Cosine SimilarityCosine similarity measures how alike two embeddings are by the angle between them — the standard relevance score behind semantic search and RAG retrieval.
- Hybrid SearchHybrid search runs keyword (BM25) and semantic (vector) retrieval together and merges the results — catching both exact terms and paraphrases.