Reranker Model

What is a reranker model?

A search usually starts with a "rough pass" that pulls in a big batch of candidates — dozens or hundreds. A reranker model's job is to look those over carefully, rescore them by how relevant they really are, and push the best one to the top.

Why add this extra step?

The first pass is fast but imprecise
Vector search is quick and can recall a lot, but it only checks "broadly similar," so it misses details and misjudges importance.
The fine pass is slow but accurate
A reranker compares the query and each candidate line by line — far more careful, but slower, so it only runs on a few candidates.
Fast and accurate together
"Rough first, fine second" — the rough pass handles speed, the fine pass handles precision. It's the standard play.

How does it judge relevance?

Score each one
It pairs your query with each candidate and outputs a relevance score.
Understands meaning, not just keywords
It matches different phrasings of the same idea, and it spots results that share words but aren't actually relevant.
Respects context
In a multi-turn chat, it uses your earlier questions to decide which candidate you actually need right now.

Where does it make things better?

Enterprise knowledge bases, e-commerce search, legal document retrieval, RAG assistants — any time the top result isn't the right one, a reranker often fixes it on the spot: hit rate goes up, and the model sees cleaner material.

Bottom line: a reranker is the "fine picking" stage of retrieval — the rough pass gives you speed, the reranker gives you precision, putting the right result first.