🔍 Cross-Encoder Re-ranking
Disabled
💡
Two-Stage Retrieval Pipeline:
Stage 1 (Bi-Encoder): Fast initial retrieval of candidate documents
Stage 2 (Cross-Encoder): Accurate re-ranking by analyzing query+document pairs together

Cross-encoders see the full context of both query and document, catching nuances that bi-encoders miss. This can significantly improve result quality at the cost of slightly slower search.
Model: cross-encoder/ms-marco-MiniLM-L-6-v2
Enable Cross-Encoder Re-ranking
Use a cross-encoder to re-rank search results for higher accuracy
Initial Candidates (Top-K) 20
Number of candidates to retrieve with bi-encoder before re-ranking
Final Results 5
Number of results to return after cross-encoder re-ranking
Model will load on first use
🚀 Coming Soon
More advanced features are planned:
  • Semantic Query Caching - Cache responses for similar queries
  • HyDE (Hypothetical Document Embeddings) - Generate hypothetical answers for better retrieval
  • Matryoshka Embeddings - Variable-dimension embeddings for speed/quality tradeoffs
  • Entity-Aware Knowledge Graph - Multi-hop reasoning across entities