Where BM25 wins, where dense wins, where you need both.

The same 60-chunk corpus from Week 5, retrieved three ways: pure BM25 lexical search, pure dense embedding search, and hybrid fusion. Pick a query and see how the rankings differ. Some queries reward exact matches (BM25 wins), some reward semantic generalization (dense wins), and most production queries — the messy mixed kind — reward the fusion of both. The fusion mode toggles between weighted sum and Reciprocal Rank Fusion.