Principle:FlagOpen FlagEmbedding Retrieval Result Reranking
Overview
A second-stage retrieval technique that re-scores candidate passages from an initial retrieval using a cross-encoder reranker to improve ranking quality.
Description
After dense retrieval returns top-k candidates, reranking re-scores each (query, passage) pair with a cross-encoder. EvalReranker truncates initial results to rerank_top_k, creates sentence pairs, calls reranker.compute_score(), and produces re-ranked results. This is computationally more expensive than embedding similarity but captures fine-grained query-passage interactions.
Usage
As the second stage of evaluation when both an embedder and reranker are configured.
Theoretical Basis
Two-stage retrieval balances efficiency (fast ANN retrieval) with accuracy (slow but precise cross-encoder scoring). Reranking only the top candidates keeps compute tractable.