What can each retrieval method miss?
Keyword search can find codes, names, and exact phrases but miss paraphrases. Vector search can bridge wording differences but overlook rare identifiers or blur related concepts. A business query may need both signals.
How are the two result sets combined?
The system retrieves candidates from lexical and vector indexes, removes duplicates, and fuses ranks or calibrated scores. A reranker may reassess the merged set. Raw scores usually cannot be added because their scales differ. Both paths must enforce document versions and access rules. Multimodal retrieval and analytics illustrates retrieval feeding an analytics workflow.
When is the extra cost justified?
Use real questions to see whether one method misses correct evidence. Hybrid search is especially useful when users mix exact identifiers, specialist terms, and natural wording. Compare recall and answer quality with indexing cost and latency.
FAQ
Is hybrid search reranking?
No. It combines recall sources; reranking reorders candidates afterward.
Can scores simply be added?
Usually not; normalize, calibrate, or fuse ranks.
When should it be tried?
When a single method demonstrably misses answers across mixed query types.