Reranking

Reorder candidates by relevance

Retrieval happens in two stages for a reason. The first must be fast enough to run across everything you have, so it is necessarily crude. The second only sees a shortlist, so it can afford to be careful.

A reranker scores each candidate against the query properly — reading them together rather than comparing precomputed vectors — and reorders accordingly.

The gain is precision where it matters. Only a handful of documents will fit in the model's context, so what sits in positions one to three decides the answer. Reranking is also the cheapest place to remove near-misses, which otherwise arrive as confident, irrelevant evidence.