# Corrective Retrieval Augmented Generation

**Authors:** Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, Zhen-Hua Ling  
**arXiv:** [2401.15884](https://arxiv.org/abs/2401.15884)  
**Date:** 2024-01-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.15884v3/

## Key points

- Hallucinations persist in large language models (LLMs) due to limitations in ensuring text accuracy based on parametric knowledge alone
- Retrieval-augmented generation (RAG) is a potential solution to complement LLMs, but its effectiveness depends on the quality of retrieved documents
- Corrective Retrieval Augmented Generation (CRAG) aims to enhance text generation robustness by evaluating and improving the quality of retrieved documents
- CRAG integrates large-scale web searches to enhance retrieval quality and uses a decompose-then-recompose algorithm to focus on key information while filtering out irrelevant details
- Experimental evaluations show significant performance improvements with CRAG compared to traditional RAG-based methods

## Summary

, , , , 

In the realm of large language models (LLMs), the issue of hallucinations persists due to inherent limitations in ensuring the accuracy of generated texts solely based on their encapsulated parametric knowledge. While retrieval-augmented generation (RAG) presents itself as a viable solution to complement LLMs, its effectiveness heavily relies on the relevance and quality of retrieved documents. This raises concerns about potential model behavior deviations in cases where retrieval fails. To address this challenge, a novel approach known as Corrective Retrieval Augmented Generation (CRAG) is introduced with the primary objective of enhancing the robustness of text generation processes. The core concept behind CRAG involves implementing a lightweight retrieval evaluator specifically designed to assess the overall quality of retrieved documents for a given query. This evaluation process results in the generation of a confidence degree, which subsequently triggers various knowledge retrieval actions based on its value. Recognizing that traditional retrieval methods from static and limited corpora may yield suboptimal results, CRAG integrates large-scale web searches as an extension to augment and enhance the quality of retrieved information. Furthermore, CRAG incorporates a unique decompose-then-recompose algorithm that enables selective focus on key information within retrieved documents while filtering out irrelevant or extraneous details. This strategic approach ensures that only pertinent data is utilized in the generation process, thereby improving overall output quality. Notably, CRAG is designed to be seamlessly integrated with diverse RAG-based approaches, offering flexibility and adaptability across different applications. Experimental evaluations conducted across four datasets encompassing both short- and long-form generation tasks demonstrate significant performance improvements achieved by CRAG when compared to conventional RAG-based methodologies. By effectively addressing issues related to retrieval accuracy and document relevance, CRAG emerges as a promising solution for enhancing text generation capabilities within large language models.