Reranking
Re-score search results using cross-encoder models for improved relevance ranking
Why do anything?
Bi-encoder search is fast but less accurate than cross-encoders. Initial results need refinement.
Why now?
Cross-encoder reranking provides significant relevance improvements on retrieved candidates.
Why this feature?
Cross-encoder model re-scores top candidates for optimal ranking. Supports multiple reranking models.
How It Works
Reranking uses cross-encoder models to refine initial retrieval results.
1
Candidate Selection
Select top-k candidates for reranking
2
Cross-Encoding
Score each query-document pair with cross-encoder
3
Re-ordering
Sort by cross-encoder scores
Why This Approach
Cross-encoders consider query-document interaction directly, providing more accurate relevance scores than bi-encoders.
Where This Is Used
Integration
retriever = client.retrievers.create(stages=[{"type": "rerank", "top_k": 20}])