Reranker
Rescores a shortlist of search results by reading the query and each result together, for better ordering.
What it is
A reranker sits after a fast first search. It reads the query alongside each candidate document and gives a relevance score, so the best few results rise to the top. In RAG systems this is one of the cheapest ways to improve answer quality.
How it works
Most are cross-encoders: the query and a document are fed into one transformer together, and it outputs a single relevance score. That is more accurate than comparing separate embeddings but too slow for millions of documents, so it only runs on the top tens or hundreds.
What it is good for
- Improving RAG retrieval
- Ordering search results
- Picking the best answer from candidates
Limits to know
- Adds latency for every query
- Only reorders; it cannot find what the first search missed
