Cross-Encoder

A cross-encoder is a transformer architecture that takes a query and a candidate document as a single joint input — typically concatenated with a separator token — and outputs one scalar relevance score. Because attention runs across both sides together, a cross-encoder can model fine-grained interactions between query terms and document terms — negation, exact matches inside a paraphrase, positional constraints — that a bi-encoder cannot. The cost is quadratic attention over the combined length and one full forward pass per (query, document) pair, which makes cross-encoders impossibly slow to run over an entire corpus at query time. They are the standard reranker architecture: a fast first-stage retriever (BM25, bi-encoder, or both) narrows the corpus to the top 50-200 candidates, and a cross-encoder rescores only those to produce the final ordering.