Hybrid Search
Combining lexical and semantic search to get keyword precision with semantic recall.
Hybrid search is a retrieval approach that combines two or more search paradigms — most commonly lexical keyword search and semantic vector search — into a single ranking, so the result set benefits from the strengths of each. The pattern has become the default production setup for retrieval-augmented generation, e-commerce search, and enterprise knowledge bases, where neither pure keyword matching nor pure embedding similarity is reliable on its own.
Mechanically, hybrid search runs the query through both pipelines in parallel and merges the two ranked lists into a single score per document. The fusion step is most often reciprocal rank fusion, a parameter-free combiner that sums the reciprocal of each document's rank in each list, but learned cross-encoders and linear-combination scorers are also common. Hybrid search also accommodates other signals — BM25, dense embeddings, sparse learned vectors like SPLADE, and metadata filters — and the architectural choice is which signals to compute, which index to query, and how to weight them.
The advantage of hybrid search is robustness: lexical matching catches exact terms, identifiers, and rare proper nouns that embeddings blur together, while semantic search catches paraphrases and conceptually related documents that lexical search would miss. The cost is operational complexity — two indexes, two query paths, and a fusion step — and a tuning problem: the weights between lexical and semantic scores often need re-balancing as the corpus or query distribution changes. Many teams also run a cross-encoder re-ranker on top, which adds a third stage of latency in exchange for precision.
Open questions include how to make the fusion stage adaptive — choosing weights per query rather than globally — and how to evaluate hybrid systems when the two pipelines fail in different ways. It is also not settled whether, in any given domain, the marginal improvement of adding a second signal is worth the indexing cost; some teams find dense retrieval alone sufficient, while others find that even a small lexical component eliminates embarrassing misses. The line between hybrid search and full multi-stage retrieval pipelines continues to dissolve.