Definition
Reranking is a post-retrieval ranking step that reorders candidates using a new estimate of their relevance to a query. The first stage searches a large corpus for plausible candidates. The reranker works only on that smaller result set.
Simple example
Vector search returns 30 documentation passages for “How should retries handle a timed-out write?” A reranker evaluates the query alongside each passage and moves passages about uncertain side effects above generic retry guidance. The application then sends only the top few passages to the model.
Why it matters
Running a costly ranking method across an entire corpus is usually impractical. The first stage can prioritize efficient, high-recall retrieval, while reranking spends more computation on the smaller result set. Better ordering can reduce irrelevant context in the passages sent to the model.
One important nuance
A reranker cannot recover a relevant item that the first stage never returned. A larger candidate pool may contain more relevant items, but it also increases reranking work and can raise latency or cost. Evaluate first-stage recall and the final ordering separately on labeled queries. That helps distinguish missing evidence from evidence that was retrieved but ranked too low.