Definition
Re-ranking
A second retrieval pass in which a more precise model re-orders the top search results before they reach the language model.
First-stage retrieval is optimized for speed across millions of chunks, so its ordering is approximate. A re-ranker, typically a cross-encoder or a hosted service like Cohere Rerank, scores each of the top 20–50 candidates against the question more carefully.
Putting the best few passages first matters because language models pay most attention to the most relevant context, and because sending fewer, better passages cuts cost and latency.
Related services
Related terms