Reranking

What is reranking?

In the first pass, a retrieval system grabs candidates with a fast but rough algorithm — like a first audition that keeps a few dozen people. Reranking brings in a stricter judge to re-examine that shortlist, order it properly, and push what you actually want to the top.

Why do it in two steps?

One precise pass is too expensive
High-accuracy models are slow and costly — you can't run them over every record in the database.
Coarse first, fine second is cheaper
Use a cheap method to pull a rough shortlist, then compute carefully on just that handful. Best of both worlds.

What does a reranker score on?

Cross-encoders
They feed the question and the candidate passage in together, judging how closely they fit.
Multiple signals
Beyond semantic similarity, a reranker can weigh freshness, authority, clicks and more.

What difference does it make?

Reranking is one of the highest value-for-money steps in RAG. It doesn't change what was recalled, but it moves the most relevant answer up a few slots — and when the AI writes its answer, it leans hardest on those first few results.

Bottom line: reranking is a second, stricter round of judging that lifts the best answer to the top of the list.