Reference · Glossary

Cross-encoder

A model that **scores query–document pairs together** — slower than bi-encoders but sharper for re-ranking top candidates.

#When to use

Re-ranking 20–50 hybrid retrieval hits before the LLM context window.

First-stage retrieval over millions of docs — use vector index instead.

Merge keyword + vector top-20 → cross-encoder scores each pair → keep top-5 for RAG.