Why rank candidates a second time?

Fast retrieval prioritizes speed over fine distinctions. It may return a passage that shares words with a query ahead of one that actually answers it. Reranking evaluates the smaller candidate set more carefully, often considering the query and each passage together.

How does a reranker assess results?

A first stage obtains candidates with keyword, vector, or hybrid search. A cross encoder or other ranking model scores each query passage pair, then sorts the set. The application may apply access rules, diversity constraints, and source checks before selecting evidence. Multimodal retrieval and analytics illustrates a broader search pipeline.

Can it recover a document that was never retrieved?

No. A reranker only reorders the candidates it receives. If the first stage omitted the answer, improve indexing, query formulation, or recall. More elaborate ranking also adds latency and cost, so evaluate the gain on representative questions.

FAQ

Is a larger reranker always better?

No. Compare relevance gains with latency and cost.

Can every document be reranked directly?

Usually not at large scale; first narrow the corpus.

How is it evaluated?

Measure top result relevance, final answer quality, latency, and first stage recall.