# Graph-Guided Concept Selection for Efficient Retrieval-Augmented Generation

**Authors:** Ziyu Liu, Yijing Liu, Jianfei Yuan, Minzhi Yan, Le Yue, Honghui Xiong, Yi Yang  
**arXiv:** [2510.24120](https://arxiv.org/abs/2510.24120)  
**Date:** 2025-10-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.24120v1/

## Key points

- The Graph-Based RAG framework enhances retrieval and question answering in Large Language Model (LLM) systems by constructing knowledge graphs (KG) from text chunks.
- G2ConS is a new approach that optimizes KG construction costs while maintaining retrieval effectiveness and answering quality.
- G2ConS incorporates a chunk selection method to reduce overall cost of KG construction and an LLM-independent concept graph to fill knowledge gaps without additional costs.
- G2ConS outperforms existing methods like GraphRAG, HippoRAG, LightRAG, KAG, FastRAG, and GraphReader in terms of construction cost efficiency, retrieval effectiveness, and answering quality across multiple real-world datasets.
- G2ConS emphasizes concept selection in graph construction to achieve consistent improvements in both cost efficiency and performance compared to traditional methods.

## Summary

, , , , 

The Graph-Based RAG framework has been instrumental in enhancing retrieval and question answering in Large Language Model (LLM) systems by constructing knowledge graphs (KG) from text chunks. This approach has proven particularly beneficial in domains like biomedicine, law, and political science, where effective retrieval often requires multi-hop reasoning over proprietary documents. However, the reliance on numerous LLM calls for entity and relation extraction from text chunks can result in prohibitive costs at scale. To address this challenge, a new approach called Graph-Guided Concept Selection (G2ConS) has been proposed. G2ConS incorporates a chunk selection method and an LLM-independent concept graph to optimize KG construction costs while maintaining retrieval effectiveness and answering quality. The chunk selection method identifies salient document chunks to reduce the overall cost of KG construction, while the concept graph helps fill knowledge gaps introduced by chunk selection without additional costs. In comparison to existing methods such as GraphRAG Edge et al. (2024), HippoRAG Jimenez Gutierrez et al. (2024), LightRAG Guo et al. (2024), KAG Liang et al. (2024), FastRAG Abane et al. (2024), and GraphReader Li et al. (2024b), G2ConS demonstrates superior performance in terms of construction cost efficiency, retrieval effectiveness, and answering quality across multiple real-world datasets. By combining KG and concept graphs in a hybrid retrieval strategy, G2ConS offers optimal performance while remaining compatible with mainstream GraphRAG approaches. Furthermore, previous efforts to enhance RAG performance on multi-hop reasoning tasks through KG construction have faced challenges due to high construction costs. Approaches like LightRAG Guo et al. (2024) and HiRAG Huang et al. (2025a) have attempted to simplify KG construction processes but may suffer from reduced accuracy on complex tasks. In contrast, G2ConS emphasizes concept selection in graph construction to achieve consistent improvements in both cost efficiency and performance compared to traditional methods. Overall, the introduction of G2ConS represents a significant advancement in optimizing KG-based RAG frameworks for efficient retrieval and question answering across diverse domains while mitigating prohibitive costs associated with large-scale operations.