Fetching from the wire…
Skills2026-03-20 · source-backed
Tuning the sparse weight to sparse_boost=1.2 in BM25+SPLADE hybrid retrieval — slight preference to keyword matches without overriding semantic coverage — yields 18.5% MRR improvement on domain corpora with exact terminology (SKUs, legal statutes, error codes). Vector generation dominates latency (>93%); query-time sparse/dense tuning is nearly free. VectorHub
Each link below shares sources, entities, or timing with this story.
Shared entities / Same source / What happened next / Tension
Both cover BM25, Vector; cite the same source (VectorHub); picks up the BM25 thread on 2026-03-22.
Shared entity: BM25 / Shared topic / What happened next / Tension
Both cover BM25; overlapping topics (bm25, dense); picks up the BM25 thread on 2026-06-08.
Shared entity: BM25 / Shared topic / What happened next
Both cover BM25; overlapping topics (bm25, corpora, dense); picks up the BM25 thread on 2026-07-31.
Shared entities / What happened next
Both cover BM25, Vector; picks up the BM25 thread on 2026-05-20.
Shared topic / Tension
Overlapping topics (bm25, dense, exact, gain, hybrid); pushes against this story (against).
Shared entity: BM25 / Shared topic / What happened next
Both cover BM25; overlapping topics (bm25, free); picks up the BM25 thread on 2026-08-07.
Both cover BM25; overlapping topics (exact, hybrid); picks up the BM25 thread on 2026-07-08.
Both cover BM25; overlapping topics (bm25, dense); picks up the BM25 thread on 2026-06-24.