Reranking

Advanced RAG has introduced mechanisms such as query rewriting, hybrid retrieval, and reranking to improve retrieval accuracy.

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.

Reranking is typically implemented using cross-encoders, which concatenate the query and candidate documents before feeding them in together; it usually reranks only the Top-50 to Top-100 candidates and finally selects Top-3 to Top-5 to pass to the generative model.

重排序（Re-ranking）机制在初步检索后引入交叉编码器（Cross-Encoder）对候选文档重新打分，提升输入大模型的上下文质量

Multi-channel retrieval simultaneously fetches candidate results through multiple different retrieval channels and then performs fusion ranking to maximize recall rate.

Rerank models are usually cross-encoders, with higher accuracy than bi-encoders but high computational cost, suitable for finely ranking a small number of candidates such as the top-50.

Reranking means that after candidate documents are obtained through vector recall, a cross-encoder model re-scores and sorts the results to improve final relevance.

The reranking model performs finer relevance scoring on candidate chunks after initial recall, effectively filtering out noise that is semantically similar but actually irrelevant.

A RAG pipeline typically includes several stages: document slicing, embedding, vector retrieval, reranking, and context assembly.

Traditional cross-encoder reranking suffers from a signal dilution effect in long-document scenarios: irrelevant content lowers relevance judgments that should have been high.

In RAG practice, mechanisms such as reranking, citation tracing and refusal to answer when there is no answer are usually needed to improve stability.

Hybrid search plus reranking is the upgrade every team pursuing better RAG should prioritize, rather than chasing fancier agent architectures.

Reranking can improve the order of results, but it cannot recover evidence that was never returned in the initial retrieval stage.

Semantic retrieval and keyword retrieval are often used together in mature RAG systems (hybrid retrieval), then results are optimized through a reranking model, to balance recall and precision.

The essence of reranking is to rebalance recall rate and precision rate, serving as a bridge component between retrieval and generation.

Before multi-path retrieval results enter reranking, they typically need to undergo deduplication and score normalization; otherwise, inconsistent score scales across different retrieval paths will introduce bias.

Ranking models perform fine-grained sorting on top of initially recalled results, pushing the most relevant content to the front.

Increasing re-ranking intensity or optimizing chunking strategies cannot compensate for the absence of the underlying data structure.

It is predicted that reranking, information extraction, and content moderation are the next batch of LLM-facing workloads to be "unbundled", with specialized small models replacing general-purpose large models and forming price competition.