Reranking

A second-pass relevance scoring step in a retrieval pipeline that reorders candidate documents after the initial vector search. The reranker (a cross-encoder model) reads each query-document pair together and assigns a more accurate relevance score than embedding similarity alone. Reranking typically improves RAG answer quality by 10–25% with minimal latency cost, making it a standard component in production retrieval pipelines.