# FunnelRAG: A Coarse-to-Fine Progressive Retrieval Paradigm for RAG

**Authors:** Xinping Zhao, Yan Zhong, Zetian Sun, Xinshuo Hu, Zhenyu Liu, Dongfang Li, Baotian Hu, Min Zhang  
**arXiv:** [2410.10293](https://arxiv.org/abs/2410.10293)  
**Date:** 2024-10-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.10293v1/

## Key points

- Large Language Models: Retrieval-Augmented Generation (RAG) is a dominant approach
- RAG methodology involves two key components: retrieval and generation
- Generators heavily rely on the effectiveness and efficiency of retrievers
- Existing retrieval paradigms have flat design, limiting performance ceiling
- FunnelRAG introduces progressive retrieval pipeline with coarse-to-fine granularity strategy
- FunnelRAG aims to distribute workload, enhance overall retrieval performance capabilities
- Extensive experiments show FunnelRAG achieves comparable metrics with 40% reduced time overhead

## Summary

In the realm of Large Language Models, Retrieval-Augmented Generation (RAG) has emerged as a dominant approach. This methodology involves two key components: retrieval and generation. The retrievers, or retrieval modules, are tasked with sourcing pertinent information to aid the generation modules, also known as generators. The performance of these generators is heavily reliant on the effectiveness and efficiency of the retrievers. However, a common issue with existing retrieval paradigms is their flat design, which treats retrieval procedures as singular events with consistent granularity. This flat retrieval approach presents two significant limitations. Firstly, it places a substantial burden on individual retrievers due to the lack of distribution of workload. Secondly, the constant granularity employed restricts the potential ceiling of retrieval performance. To address these shortcomings, a new progressive retrieval paradigm has been proposed for RAG - FunnelRAG. FunnelRAG introduces a progressive retrieval pipeline that incorporates a coarse-to-fine granularity strategy. By leveraging varying levels of granularity, quantity scaling from large to small, and capacity ranging from low to high, FunnelRAG aims to alleviate the burden on individual retrievers while simultaneously enhancing overall retrieval performance capabilities. Extensive experiments conducted on FunnelRAG have demonstrated its ability to achieve comparable retrieval performance metrics while reducing time overhead by nearly 40 percent. This innovative approach represents a significant advancement in balancing effectiveness and efficiency within Retrieval-Augmented Generation systems.