LLM concepts

Reranking

A retrieval step that reorders initially retrieved results using a more precise scoring model or rule before passing context to the next stage.

Reranking improves retrieval quality by separating candidate generation from final selection. A fast vector or keyword search finds a broad set of possible matches; a reranker then scores those candidates more carefully against the query.

This is useful when the first-stage retriever returns semantically related but not actually useful passages. The cost is extra latency and compute, so reranking should be measured against downstream answer quality, not used automatically.