# RQ-RAG: Learning to Refine Queries for Retrieval Augmented Generation

**Authors:** Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, Jie Fu  
**arXiv:** [2404.00610](https://arxiv.org/abs/2404.00610)  
**Date:** 2024-03-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.00610v1/

## Key points

- Authors introduce Retrieval-Augmented Generation (RAG) to enhance accuracy by integrating external documents into response generation process
- Existing RAG implementations may overlook complexities in queries, requiring further clarification or decomposition for precise responses
- RQ-RAG proposed as a solution with explicit rewriting, decomposition, and disambiguation capabilities
- Experimental evaluations using a 7B Llama2 model show RQ-RAG outperforms previous state-of-the-art models by an average of 1.9% across three single-hop question-answering datasets
- RQ-RAG exhibits enhanced performance in handling complex multi-hop tasks
- Authors provide access to code on GitHub at https://github.com/chanchimin/RQ-RAG for further exploration and implementation

## Summary

, , , , 

In their paper titled "RQ-RAG: Learning to Refine Queries for Retrieval Augmented Generation," authors Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu delve into the challenges faced by Large Language Models (LLMs) in generating accurate responses. The authors introduce Retrieval-Augmented Generation (RAG) as a methodology that integrates external documents into the response generation process to enhance accuracy. However, existing RAG implementations may overlook complexities in queries that require further clarification or decomposition for precise responses. To address this limitation, the authors propose RQ-RAG as a solution with explicit rewriting, decomposition, and disambiguation capabilities. Experimental evaluations using a 7B Llama2 model show that RQ-RAG outperforms previous state-of-the-art models by an average of 1.9% across three single-hop question-answering datasets and exhibits enhanced performance in handling complex multi-hop tasks. The authors provide access to their code on GitHub at https://github.com/chanchimin/RQ-RAG for further exploration and implementation of their proposed methodology.