Reranking
Also known as: reranker · cross-encoder
Reranking is a second, more careful pass over the chunks that retrieval found. A reranker (often a cross-encoder) scores each (query, chunk) pair and reorders them so the most relevant come first.
It’s cheap and high-impact:
- First pass (vector search): fast, approximate, returns ~top 20.
- Rerank: slower, precise, reorders to a tight top 5.
The easy win: adding a reranker is one of the highest-ROI upgrades to a RAG pipeline. Same model, same data, better-ordered context → better answers.