# 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/2401.15884v2/

## Key points

- Large language models (LLMs) exhibit hallucinations due to limitations in accuracy based on parametric knowledge
- Retrieval-augmented generation (RAG) proposed as a solution to complement LLMs
- Concerns about RAG effectiveness if retrieval fails due to relevance of retrieved documents
- Corrective Retrieval Augmented Generation (CRAG) introduced by researchers including Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling to enhance text generation robustness
- CRAG incorporates a lightweight retrieval evaluator for assessing quality of retrieved documents and guiding knowledge retrieval actions
- CRAG leverages large-scale web searches to augment retrieval process and focuses on key information while filtering out irrelevant content
- Plug-and-play compatibility with various RAG-based approaches for seamless integration into existing models
- Experimental evaluations show significant performance improvements with CRAG in mitigating hallucinations and enhancing text generation quality

## Summary

, , , , 

Large language models (LLMs) have shown to exhibit hallucinations due to the limitations in securing the accuracy of generated texts solely based on their parametric knowledge. While retrieval-augmented generation (RAG) has been proposed as a viable solution to complement LLMs, its effectiveness heavily relies on the relevance of retrieved documents, leading to concerns about model behavior if retrieval fails. In response to this challenge, a team of researchers including Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling introduces Corrective Retrieval Augmented Generation (CRAG) as a novel approach to enhance the robustness of text generation. CRAG incorporates a lightweight retrieval evaluator designed to assess the quality of retrieved documents for a given query. This evaluator provides a confidence degree that guides different knowledge retrieval actions based on the assessed relevance. Recognizing that traditional retrieval from static and limited corpora may yield sub-optimal results, CRAG leverages large-scale web searches as an extension to augment the retrieval process. Additionally, CRAG implements a decompose-then-recompose algorithm that selectively focuses on key information within retrieved documents while filtering out irrelevant content. One notable feature of CRAG is its plug-and-play compatibility with various RAG-based approaches, allowing seamless integration into existing models. Experimental evaluations conducted across four datasets covering both short- and long-form generation tasks demonstrate significant performance improvements when employing CRAG. The results highlight CRAG's ability to mitigate hallucinations in generated texts and enhance overall text generation quality through its corrective retrieval mechanisms. In conclusion, Corrective Retrieval Augmented Generation presents a promising advancement in addressing the challenges associated with hallucinations in large language models by improving the reliability and robustness of text generation processes through enhanced retrieval strategies and selective content focus.