GlossaryRanking
Reranker
A second-stage model that re-orders retrieval results by relevance to the original query.
A reranker sits between the first-stage retriever (which is fast but coarse) and the LLM. It takes the top N candidates from the retriever and scores each one against the query using a heavier model - typically a cross-encoder. Rerankers dramatically improve precision at the cost of latency. Cohere's Rerank, Jina Reranker, and BGE-Reranker are commonly used.