# Reranking Models

> Cross-encoder models that rerank initial retrieval results for improved relevance. More accurate than bi-encoders but slower, typically applied to top-k candidates.

- **URL:** https://www.pinecone.io/learn/series/rag/rerankers/
- **Category:** Machine Learning Models
- **Tags:** Reranking, cross-encoder, Rag
- **Updated:** 2026-03-11 12:14
- **Canonical page:** https://vectordb.ever.works/ko/items/reranking-models

## Details

## 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

---

_This Markdown mirror is generated by Ever Works for AI agents. The canonical HTML page is at https://vectordb.ever.works/ko/items/reranking-models._