Cross-Encoder

The reranking stage typically uses a Cross-Encoder, which jointly encodes the query and document by concatenating them, enabling more accurate relevance judgment than bi-encoder models but at a higher computational cost.

Reranking is the step of performing a secondary sort on retrieval results to improve relevance.

Cross-encoders are more accurate than bi-encoders because they concatenate [query, doc] and feed it into a Transformer, letting the attention mechanism compare directly at the word level, but they cannot pre-build an index and can only be used for fine ranking.

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.

Rerankers are typically implemented using cross-encoders, which concatenate the query and document and feed them jointly into a Transformer model, achieving significantly higher ranking precision than the initial retrieval stage

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.

A Cross-Encoder concatenates a query with each candidate document and feeds them jointly into the same model, capturing word-level interactions through the attention mechanism; its ranking precision is superior to vector similarity, but latency grows linearly with the size of the candidate set.

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

Reranking 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.

重排（Reranking）通常使用交叉编码器对每个片段与用户问题的相关性重新评估

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.

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

After softmax normalization, the logits output by the cross-encoder have an exponential amplifying effect on citation probability, so tiny differences in vector similarity are insufficient to decide the outcome.

In actual measurements, the difference in dense vector similarity may be only 0.10 to 0.15, but the cross-encoder's score for text chunks containing engineering triples can jump by 4.0 to 6.0 logits.

The hybrid RAG pipeline of modern generative search engines typically first recalls 20 candidates with a bi-encoder, then re-ranks them precisely with a cross-encoder, while graph matching provides deterministic triples.

A cross-encoder concatenates the query and document and feeds them together into a Transformer, which can capture deeper relevance signals, but document representations cannot be precomputed, and each query-document pair requires one forward pass, so it is only suitable for reranking a small number of candidates.

The standard approach is to use the cross-encoder only to rerank the dozens of candidates returned by retrieval, forming a two-stage pipeline of "coarse recall + fine ranking".

Cross-encoders cannot precompute, and each (query, document) pair requires one forward inference, which is infeasible for million-scale corpora.

A cross-encoder takes roughly 20-50 milliseconds to evaluate the top 25 candidates, and introducing a reranker typically reduces irrelevant context injection by 40%-60%.