Hybrid retrieval & Reciprocal Rank Fusion

Fusion lab: two arms over the same forty-two records. The semantic arm puts 17 of 31 answers first and the lexical arm 21, but each is blind where the other sees — 0 of 6 bare codes first against 6, 9 of 13 paraphrases against 11. Fused by rank at k=60 they reach 24 first places and MRR 0.8710, without moving recall@10 from its 1.0000, because the win is order and not coverage. Then the bill: five questions drop the answer below its best arm, worst by 7 places, because a row both arms returned outranks a row only one loved. The mean cosine gap is 0.0826 and the cheapest BM25 term 2.5639, 31x bigger, which is why scores cannot be added.

This Memory Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.