Cross-Encoder
A model that jointly encodes a query and candidate to produce a high-quality relevance score at higher computational cost than independent embeddings.
Information-Retrieval Context
A cross-encoder processes the query and candidate together, allowing token-level interactions and full attention between both inputs. This can produce a more accurate relevance score than independent embeddings, but every query-candidate pair requires a separate inference.
Scaling Boundary
Cross-encoders are usually too expensive for first-stage retrieval across a large corpus. They fit naturally in a reranking stage after a cheaper retriever has reduced the candidate set.