Hybrid Search
Keyword search and vector search, combined and merged.
Running lexical retrieval (BM25 or full-text) alongside vector retrieval and blending the two ranked lists, so exact terms and paraphrases both work.
This is the pragmatic default for real corpora. Vector search finds the conceptually relevant passage; keyword search guarantees exact identifiers are not missed.
The reason it works is that the two methods fail in opposite directions. Vectors handle paraphrase and lose rare exact tokens; lexical search nails identifiers, error codes and product names and misses synonyms. Merging the ranked lists covers both, and needs no labelled data to set up.
The error code nobody could find
Support search cannot find documents by ticket or error identifiers.
Embeddings place a string like ERR-4021 essentially at random. Keyword search finds it instantly. Running both and fusing the ranks costs little and fixes the entire class of exact-token lookups.
- blogContextual retrieval in AI systems ↗
The practical hybrid-retrieval article: chunking, embeddings, BM25, rank fusion and reranking, plus prepending chunk context to reduce retrieval failures.
- paperDense passage retrieval for open-domain question answering ↗
Dual-encoder dense retrieval trained on questions and passages, and the comparison to BM25 that started the dense-versus-sparse debate.
Semantic Search
Search by meaning instead of matching words.
Reranking · Reranker
A second pass that reorders candidates by actual relevance.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Vector Database · Vector store
Storage built for finding the nearest vectors, fast.