# A Hybrid RAG System with Comprehensive Enhancement on Complex Reasoning

**Authors:** Ye Yuan, Chengwu Liu, Jingyang Yuan, Gongbo Sun, Siqi Li, Ming Zhang  
**arXiv:** [2408.05141](https://arxiv.org/abs/2408.05141)  
**Date:** 2024-08-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.05141v1/

## Key points

- Authors Ye Yuan, Chengwu Liu, Jingyang Yuan, Gongbo Sun, Siqi Li, and Ming Zhang introduce a retrieval-augmented generation (RAG) framework to improve accuracy and reduce hallucinations in large language models (LLMs) by integrating external knowledge bases.
- The hybrid RAG system presented in the study is enhanced through optimizations aimed at enhancing retrieval quality, augmenting reasoning capabilities, and refining numerical computation ability.
- Strategies implemented include refining text chunks and tables in web pages, adding attribute predictors to reduce hallucinations, conducting LLM Knowledge Extractor and Knowledge Graph Extractor processes, and building a reasoning strategy incorporating all references.
- Evaluation on the CRAG dataset through the Meta CRAG KDD Cup 2024 Competition showed significant enhancements in complex reasoning capabilities with improved accuracy and reduced error rates compared to baseline models.
- The technical report received the 3rd prize in Task 1 of the Meta CRAG KDD Cup 2024 competition.
- The source code for their system is publicly available at https://gitlab.aicrowd.com/shizueyy/crag-new.

## Summary

In their paper titled "A Hybrid RAG System with Comprehensive Enhancement on Complex Reasoning," authors Ye Yuan, Chengwu Liu, Jingyang Yuan, Gongbo Sun, Siqi Li, and Ming Zhang introduce a retrieval-augmented generation (RAG) framework that enables large language models (LLMs) to improve accuracy and reduce hallucinations by integrating external knowledge bases. The hybrid RAG system presented in the study is enhanced through a series of optimizations aimed at enhancing retrieval quality, augmenting reasoning capabilities, and refining numerical computation ability. The authors implemented various strategies such as refining text chunks and tables in web pages, adding attribute predictors to reduce hallucinations, conducting LLM Knowledge Extractor and Knowledge Graph Extractor processes, and building a reasoning strategy incorporating all references. The system was evaluated on the CRAG dataset through the Meta CRAG KDD Cup 2024 Competition. Both local and online evaluations demonstrated significant enhancements in complex reasoning capabilities. Local evaluations showed improved accuracy and reduced error rates compared to baseline models, resulting in notable score increases. Online assessments further validated the performance and generalization capabilities of the proposed system. The technical report by Yuan et al. received the 3rd prize in Task 1 of the Meta CRAG KDD Cup 2024 competition. The source code for their system is publicly available at https://gitlab.aicrowd.com/shizueyy/crag-new. This comprehensive study showcases how optimizing a hybrid RAG system can significantly enhance complex reasoning abilities in large language models.