Reranking

A second-stage retrieval step where a smaller cross-encoder scores each candidate document against the query, boosting precision above vector similarity alone.