Reranking
Also called Reranker, Re-ranking
A second search step that reorders a shortlist of results by how well each one actually answers the question, so the best come first.
What is reranking?
Reranking is a two-step approach to search. First, a fast search pulls a shortlist of possibly relevant passages, say the top 50. Then a slower, more careful model, called a reranker, scores each one against the question and puts them in a better order. Only the top few go on to the AI model.
It's common in RAG systems, where the quality of the passages decides the quality of the answer.
How does it work?
The first step, often semantic search or hybrid search, compares the question and each passage separately, as pre-computed embeddings. That's fast enough to scan millions of passages, but it can miss nuance.
A reranker reads the question and a passage together and judges how well the passage answers it. That's much more accurate, but too slow to run on everything, which is why it only sees the shortlist.
What is it commonly confused with?
- The first search. A reranker can't find a passage the first step missed; it only reorders what it's given. If retrieval misses the right chunk, reranking won't save it.
- The AI model itself. Rerankers are small, specialised models that output a relevance score, not text.
Why does it matter?
Adding a reranker is one of the simplest ways to improve a RAG system's answers: the model sees fewer, better passages, which improves accuracy and cuts the tokens you pay for. The cost is a little extra time on each search.
