PRODUCTION SCENARIO
A payer's claims-appeal assistant runs hybrid search over 300,000 policy documents. An audit of 200 failed answers shows the passage that answers the question is in the fused candidate set every time, but sits below three irrelevant passages in the context sent to the generator.
Which next experiment addresses this ranking problem?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Score the fused candidates with a cross-encoder reranking model
Recall is already fine and ordering is not, so the experiment is a second-stage scorer over the candidates the first stage returned - and it has to be measured, because reranking buys ordering at a real latency cost. Passing 50 candidates dilutes the context rather than ordering it. Reweighting the fusion swaps one first-stage proxy for another. Contextual chunk rewriting is a recall fix for passages that never surface, which is not this failure.