Hybrid RAG Sparse Boost Tuning: BM25 + SPLADE With sparse_boost=1.2 Delivers 18.5% MRR Gain on Domain-Specific Corpora

VectorHub by Superlinked·medium signal

Production benchmarks show that tuning the sparse weight to `sparse_boost=1.2` in BM25+SPLADE hybrid retrieval — giving a slight preference to keyword matches without overriding semantic coverage — yields an 18.5% improvement in Mean Reciprocal Rank on domain corpora with exact terminology (product SKUs, legal statutes, error codes). RRF fusion is the recommended baseline before tuning. Critically, the vector generation step dominates latency (>93%) while the actual search and fusion steps contribute <7%, so query-time sparse/dense balance tuning is nearly free compared to re-indexing.