# SimGRAG: Leveraging Similar Subgraphs for Knowledge Graphs Driven Retrieval-Augmented Generation

**Authors:** Yuzheng Cai, Zhenyue Guo, Yiwen Pei, Wanrui Bian, Weiguo Zheng  
**arXiv:** [2412.15272](https://arxiv.org/abs/2412.15272)  
**Date:** 2024-12-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.15272v1/

## Key points

- Recent advancements in large language models (LLMs) have shown versatility across tasks
- Retrieval-augmented generation (RAG) addresses hallucinations in LLMs by using external knowledge sources like knowledge graphs (KGs)
- Similar Graph Enhanced Retrieval-Augmented Generation (SimGRAG) aligns query texts with KG structures through a two-stage process
- Emphasis on plug-and-play usability and scalability in SimGRAG
- GSD metric and optimized subgraph retrieval algorithm enhance efficiency and scalability
- SimGRAG outperforms existing methods in question answering and fact verification tasks
- Performance of SimGRAG depends on the capabilities of underlying LLMs, which may affect performance for complex queries

## Summary

, , , , 

Recent advancements in large language models (LLMs) have demonstrated remarkable versatility across various tasks. To address the issue of hallucinations in LLMs, retrieval-augmented generation (RAG) has emerged as a powerful approach, utilizing external knowledge sources like knowledge graphs (KGs). In this study, we delve into the realm of KG-driven RAG and introduce a novel method known as Similar Graph Enhanced Retrieval-Augmented Generation (SimGRAG). The core focus of SimGRAG lies in effectively aligning query texts with KG structures through a two-stage process. Firstly, the query-to-pattern alignment involves using an LLM to transform queries into a desired graph pattern. Secondly, the pattern-to-subgraph alignment quantifies the alignment between the pattern and candidate subgraphs using a graph semantic distance (GSD) metric. Additionally, an optimized retrieval algorithm has been developed to efficiently identify the top-k subgraphs within 1-second latency on a 10-million-scale KG. One key aspect of SimGRAG is its emphasis on plug-and-play usability and scalability. By leveraging the query-to-pattern and pattern-to-subgraph alignment paradigm, SimGRAG ensures context conciseness and prevents entity leaks. The introduction of the GSD metric and the optimized subgraph retrieval algorithm further enhance retrieval efficiency and scalability. Extensive experiments conducted across different KG-driven RAG tasks have consistently shown that SimGRAG outperforms existing methods in question answering and fact verification. The method's ability to align query texts with KG structures effectively contributes to its superior performance compared to state-of-the-art approaches. However, it is important to note that the performance of SimGRAG heavily relies on the underlying capabilities of LLMs. Lower-quality or less capable LLMs may lead to degraded performance, especially in scenarios requiring advanced reasoning skills for complex queries. In conclusion, our study introduces SimGRAG as an innovative approach for KG-driven RAG that offers improved usability, context conciseness, and scalability. By combining large language models with knowledge graphs effectively, SimGRAG sets a new standard for addressing challenges in information retrieval tasks.