Why it matters
Reranking improves precision. Understanding shapes RAG design.
Advertisement
The architecture
Bi-encoder: retrieve top-50-100.
Cross-encoder: score (query, doc) pairs.
Return top-5.
Advertisement
How it works end to end
Cross-encoder more accurate but slow.
Rerank latency 50-200ms.
BGE / Cohere / ColBERT choices.