Reranking
Reranking: Reordering a limited candidate set with a more accurate model than the initial retrieval stage. The first stage favors recall and speed; the second reads query and document more carefully. Cross-encoders and LLMs can improve precision at the cost of latency.
In italiano: RerankingLegge I · Embedding, feature e spazi vettorialiThe first stage favors recall and speed; the second reads query and document more carefully. Cross-encoders and LLMs can improve precision at the cost of latency.
Examples
- Cross-encoder over the first 100 results
- Reranking before prompting
- Relevance-based ordering