In the FieldDesk industrial service scenario, a technician queries: "How to stop hydraulic fluid discharge from primary seal assembly on Model X-840 with fault code ALM-3042." Why does deploying either dense vector search or lexical BM25 search in isolation fail to retrieve the authoritative troubleshooting manual page?
Day 11: Combine Dense and Lexical Results with Hybrid Search
Combine Dense and Lexical Results with Hybrid Search
Loading lesson content...
In FieldDesk's hybrid retrieval pipeline using Reciprocal Rank Fusion (RRF) with default smoothing constant k = 60, Document A ranks 1st in dense retrieval but is missing from the top-50 lexical candidates. Document B ranks 5th in dense retrieval and 6th in lexical retrieval. How does RRF evaluate their relative priority?
When architecting the hybrid search layer for PolicyDesk, the engineering team evaluates Convex Linear Score Combination versus Reciprocal Rank Fusion (RRF). Why is RRF recommended as the primary production baseline over naive linear score addition?