Reranking Models

Overview

Reranking models refine initial retrieval results by computing precise relevance scores between queries and retrieved documents.

Architecture

Cross-Encoders

• Process query and document together
• More accurate than bi-encoders
• Expensive: Can't pre-compute
• Applied to top-k results only

Popular Models

• Cohere Rerank: Commercial API
• bge-reranker-v2-m3: BAAI multilingual reranker
• Cross-Encoder Models: HuggingFace Sentence Transformers
• Jina Reranker: Various sizes available

Typical Pipeline

1. Retrieval: Get top-100 candidates with vector search
2. Rerank: Apply reranker to refine order
3. Select: Choose top-10 for LLM context

Benefits

• Improved Precision: Better relevance ranking
• Better RAG: More relevant context for LLMs
• Hybrid Approach: Combine speed (retrieval) + accuracy (reranking)

When to Use

• High-precision requirements
• When latency allows (~100ms)
• Production RAG systems
• Search quality matters

Pricing

• Open Source: Free (bge-reranker, others)
• Cohere Rerank: Usage-based API pricing
• Self-hosted: Compute costs