Techniques & Methods
Reranker in plain English.
Also known as: reranking model,cross-encoder,rerank
The one-sentence version
A second-stage model that re-orders search results by relevance so the best passages reach the AI first.
A reranker is a model that takes a query and a shortlist of candidate documents — typically the top 20 to 100 results from a fast vector or keyword search — and scores each one for how well it actually answers the query, producing a better-ordered list. First-stage retrieval is optimised for speed over millions of items and is only roughly right; a reranker is slower but far more accurate because it reads the query and each candidate together. In a RAG pipeline this is the difference between the model seeing the three passages that matter and seeing three near-misses. Cohere, Pinecone, and several open-weight projects offer rerankers, and most serious retrieval systems now include one. The cost is extra latency and compute on every query, which is why the candidate list is kept short.