Topic 2: Rerankers

6 min read·21 Sept 2026

Reranker	How it decides	Calls per query	Typical latency	Strength
Cross-encoder	One small model run per candidate, locally	0 LLM calls	10-200 ms for 25-50 candidates	Best accuracy per millisecond
LLM pointwise	Scores each passage	1 per batch	Hundreds of ms per batch	Follows custom criteria
LLM listwise	Orders a window of passages	1 per window	Hundreds of ms per window	Strong ordering; uses comparisons
LLM pairwise	“A or B?”	Many	Highest	Most accurate per comparison; rarely affordable
Hosted API	Someone else’s cross-encoder	1 network call	50-300 ms	No model to run; pay per search

2.1 Cross-Encoder Rerankers

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.