A cross-encoder is a model that feeds the query and a candidate document jointly into a single Transformer to produce a relevance score. Unlike a bi-encoder, it doesn't compute separate vector representations, so it's much slower but much more accurate. Sentence-BERT (Reimers & Gurevych, 2019) popularized this architecture. In practice a bi-encoder handles the first retrieval pass and a cross-encoder kicks in during Reranking — the combination is widely regarded as the best speed/accuracy trade-off in modern RAG pipelines.
MEVZU N°124ISTANBULYEAR I — VOL. III
Glossary · Advanced · 2019
Cross-encoder
A transformer architecture that processes the query and a candidate document jointly to score relevance.
- EN — English term
- Cross-encoder
- TR — Turkish term
- Cross-encoder