Hybrid Search
A retrieval strategy that combines lexical and semantic signals, such as BM25 and vector similarity, to improve recall and relevance.
Information-Retrieval Context
Hybrid search combines lexical and semantic retrieval signals. Sparse methods such as BM25 preserve exact terms, identifiers, and code tokens, while embedding search captures semantic similarity; rank fusion or a learned reranker can combine their complementary evidence.
Score Boundary
Raw sparse and dense scores generally do not share a meaningful scale. Simply adding them can be unstable, so normalization, rank fusion, or an explicitly trained combination method should be evaluated.
Lexical and vector retrieval scores are produced on different scales and should not usually be added without calibration. Rank fusion, score normalization, or learned ranking should be selected using retrieval-quality measurements alongside query latency.