ML//RAG//reranker

A reranker is a second-pass model that re-scores retrieved documents for their relevance to the query.

A reranker is a second-pass model that re-scores retrieved documents for their relevance to the query.

A first pass (BM25 or embeddings) retrieves candidates cheaply, and the reranker then uses cross-attention to evaluate each query and document pair. It is much more accurate than bi-encoder similarity but too slow for a full-index search, which is why the pipeline has two stages.