Reranking Models
Reranking models are second-stage retrieval models that rescore candidate results after a cheaper recall step. In Episode 17: 向量模型工程师：AI 的隐藏瓶颈与新时代的信息迷宫, N 同学 / N Student describes a common Retrieval-Augmented Generation flow: use Vector Model Engineering to recall a batch of candidates, then use a more specialized reranker to choose the passages most relevant to the user question.
The reason is practical. First-stage recall must search many items quickly, so it can return plausible but weak matches. A reranker can inspect query-document pairs more directly, use a narrower Semantic Search Relevance definition, and improve the evidence passed to the final language model.
Key Claims
- Reranking separates fast recall from more precise relevance scoring.
- A reranker is only useful when the candidate pool contains the right answer often enough.
- Reranking quality depends on labels, hard negatives, task definition, and evaluation metrics.
- In RAG, reranking affects generation quality because the final model can only answer from what it receives.
- Agent search could benefit from better reranking because agents otherwise waste steps on mechanical keyword searches.
Connections
- Retrieval-Augmented Generation - downstream pipeline where reranking is commonly used.
- Vector Model Engineering - first-stage recall layer that rerankers refine.
- Hard Negative Mining - training mechanism for making rerankers distinguish close wrong answers.
- Semantic Search Relevance and AI Search Evaluation - standards used to judge ranking quality.
- Deep Research and Agent Harness - agentic search contexts where ranking determines useful evidence.