What is Hybrid Search?
Also called Hybrid Retrieval.
Hybrid search combines keyword matching with vector based semantic matching and merges the two result lists into one ranking. Keyword scoring catches exact terms, identifiers, and rare names, while embeddings catch paraphrase and intent. The blended result is usually more reliable than either method alone, particularly on technical corpora full of codes and product names.
The two retrieval families fail in opposite ways. Lexical scoring, typically a variant of BM25, ranks documents by how often query terms appear relative to how rare those terms are across the collection. It is exact and cheap, and it is helpless when the user's wording differs from the document's wording.
Vector retrieval tolerates paraphrase but blurs literal detail. A serial number, an error code, or a version string may barely influence a document embedding, so a system relying on vectors alone can miss the one document that names it. Running both retrievers and combining their results covers each gap with the other's strength.
Merging requires care because the two score scales are not comparable. Reciprocal rank fusion is a common solution, since it uses only the position of each result within its own list and needs no calibration. Weighted score normalization is the alternative and offers finer control at the cost of tuning per corpus.
Hybrid retrieval typically precedes a reranking step, where a stronger model scores a shortlist drawn from both retrievers. Reported gains vary by domain, and the honest summary is that hybrid helps most where vocabulary is technical and precise, and helps least where content is conversational prose with consistent terminology.
Key points
- Runs keyword and vector retrieval, then merges the lists.
- Keyword half catches codes, names, and exact terms.
- Vector half catches paraphrase and described intent.
- Reciprocal rank fusion merges without score calibration.
In practice
A developer searches for timeout error 504 on upload. The keyword retriever locks onto the literal code 504 and returns the reference page that lists it. The vector retriever returns a guide about large file uploads failing, which never mentions the code. Fusing both lists puts the reference page and the practical guide in the top three.