This survey addresses the limitations of standard Retrieval-Augmented Generation (RAG) when dealing with complex relational data. It introduces GraphRAG, a methodology that leverages structural information within knowledge graphs to enhance retrieval precision and comprehensiveness. Aimed at researchers and engineers, GraphRAG offers a path to more accurate, context-aware LLM responses by mitigating issues like hallucination and outdated information through relational knowledge integration.

How It Works

GraphRAG formalizes a three-stage workflow: Graph-Based Indexing (G-Indexing), Graph-Guided Retrieval (G-Retrieval), and Graph-Enhanced Generation (G-Generation). G-Indexing focuses on constructing high-quality graph databases. G-Retrieval tackles challenges like exponential subgraph growth and measuring similarity between textual queries and graph data, exploring various retrieval paradigms and enhancements. G-Generation integrates retrieved graph data into formats compatible with LLMs to produce superior outputs. This approach uniquely captures relational knowledge, a significant advantage over text-only RAG systems.

Quick Start & Requirements

This repository hosts a survey paper detailing GraphRAG methodologies. The primary resource is the paper itself, available at https://arxiv.org/abs/2408.08921. No installation, code, or specific system requirements are provided within this README.

Highlighted Details

• Presents the first comprehensive survey of GraphRAG methodologies, formalizing its workflow into indexing, retrieval, and generation stages.
• Categorizes core technologies, training methods, downstream tasks, application domains, evaluation metrics, and industrial use cases.
• Details challenges in graph retrieval, including explosive candidate subgraphs and the need for robust graph-text similarity measurement.
• Compiles an extensive list of related research papers, categorized by indexing type, retrieval method, and enhancement techniques.

Maintenance & Community

The survey was first released on arXiv in August 2024 and updated in September 2024 with the repository's creation. Direct contact for questions or suggestions is available via email. No community channels like Discord or Slack are listed.

Licensing & Compatibility

As this is a survey paper, no specific software license is provided. Compatibility for commercial use or closed-source linking is not applicable.

Limitations & Caveats

The survey focuses on existing GraphRAG methodologies and does not detail specific implementation limitations. It highlights inherent challenges within the GraphRAG process, such as efficient subgraph retrieval and accurate similarity measurement, and points towards future research directions.