Reranker
A reranker is a second-stage ranking component in information retrieval systems, positioned after initial retrieval to perform more refined relevance assessment and reordering of candidate results. Its core function is to compensate for ranking imprecision introduced by computational efficiency constraints in the initial retrieval stage through deep cross-attention matching between queries and candidate documents, typically using pretrained language models, thereby substantially improving the accuracy and relevance of final retrieval results at the cost of modest additional computation.
Core Facts
Rerankers typically adopt a Cross-Encoder architecture, in contrast to the Bi-Encoder architecture used in vector retrieval.
Cross-Encoder reranking models can achieve more precise result ranking at the cost of higher computational overhead.
Reranking uses a cross-encoder to score query-document pairs one by one, with accuracy far higher than vector similarity but at higher computational cost, so fine ranking is done only on the small number of candidates after coarse ranking, forming a two-stage architecture of coarse ranking plus fine ranking.
Timeline (last 90 days)
Replacing the Retriever or Reranker directly changes the context fed into the model, which in turn affects the output quality of the entire RAG system, so any component change requires evaluating whether retrieval quality has degraded.
Reranking uses a cross-encoder to score query-document pairs one by one, with accuracy far higher than vector similarity but at higher computational cost, so fine ranking is done only on the small number of candidates after coarse ranking, forming a two-stage architecture of coarse ranking plus fine ranking.
Common optimization techniques for vector retrieval accuracy include adjusting chunking strategies, using reranking models (Reranker), and hybrid retrieval (keyword + semantic).
All Facts (10)
Rerankers typically adopt a Cross-Encoder architecture, in contrast to the Bi-Encoder architecture used in vector retrieval.
75%VerifiedCross-Encoder reranking models can achieve more precise result ranking at the cost of higher computational overhead.
70%VerifiedReranking uses a cross-encoder to score query-document pairs one by one, with accuracy far higher than vector similarity but at higher computational cost, so fine ranking is done only on the small number of candidates after coarse ranking, forming a two-stage architecture of coarse ranking plus fine ranking.
65%VerifiedReranking uses a Cross-Encoder model to re-score and re-rank the candidate chunks recalled by Top-K, with precision far higher than that of a Bi-Encoder.
65%UnverifiedReranker模型（如Cohere Rerank、BGE-Reranker、bce-reranker）同时考虑Query和Document的交互信息，排序精度更高但计算成本也更大
90%UnverifiedAI Search使用两阶段检索架构：第一阶段基于向量相似度粗筛，第二阶段使用Reranker模型精排
90%Unverified在RAG实际工程中，通常先召回Top-50到Top-100的候选文档，再通过Reranker精排后取Top-5到Top-10注入上下文
75%UnverifiedDocument chunking strategy is one of the aspects with the greatest impact on the final effectiveness of RAG systems, and research indicates its influence even surpasses model selection itself.
60%UnverifiedReplacing the Retriever or Reranker directly changes the context fed into the model, which in turn affects the output quality of the entire RAG system, so any component change requires evaluating whether retrieval quality has degraded.
50%UnverifiedCommon optimization techniques for vector retrieval accuracy include adjusting chunking strategies, using reranking models (Reranker), and hybrid retrieval (keyword + semantic).
50%