🎯 Production RAG • Cohere Rerank • BGE-Reranker-v2 • ColBERTv2 • NDCG@10
RAG Reranker & MRR / NDCG Calibrator
Quantify exact retrieval precision improvements when inserting a cross-encoder reranking stage over standard dense vector cosine similarity (OpenAI text-embedding-3 / Voyage AI).
Reranker Architecture
NDCG@10 Score
0.864
+34.8% vs Bi-Encoder
MRR@10 (Rank 1 Accuracy)
0.812
+42.1% vs Bi-Encoder
Added Rerank Latency
+18 ms
p95 API response
Live Passage Reordering Simulation
Dense → Cross-Encoder| Initial Rank | Passage Content Snippet | Cosine Sim | Rerank Score | Status |
|---|
📐 Why Cross-Encoders Dominate Bi-Encoders:
Bi-encoders compute query and document representations independently ($u = E(q)$, $v = E(d)$), losing all query-token-to-document-token cross-attention interactions. Cross-encoders feed $(q, d)$ together through deep multi-head self-attention, detecting nuanced lexical relevance that cosine embeddings miss.