R · RAG · reranker

Reranker

A secondary model that reorders retrieval results for relevance.

In one sentence

A reranker is a (usually cross-encoder) model that takes the top-N candidates from a first-pass retriever and reorders them by predicting a fine-grained relevance score for each query-passage pair.

When it matters

When the first-pass retriever returns many plausible candidates and you need to spend more compute on ranking only the survivors.

A real-world example

helpcode KB retrieves 200 candidates per query (fast, fuzzy) and reranks the top 50 with a cross-encoder; the LLM only sees the top 12. End-to-end latency stays under 1.4 s.


Curated by helpcode research team · Last reviewed 2026-05-22