GraphRAG: Experiments with Graph-Based RAG Systems

prj_graphrag Project

prj_graphrag is a project for experimenting with Microsoft's GraphRAG. It ran through March 2025, with 3 commits total.

What is GraphRAG?

Unlike standard RAG, GraphRAG indexes documents as a graph structure for retrieval:

graph LR
    subgraph Standard RAG
        A1[문서] --> B1[청킹]
        B1 --> C1[벡터화]
        C1 --> D1[유사도 검색]
    end

    subgraph GraphRAG
        A2[문서] --> B2[엔티티 추출]
        B2 --> C2[관계 그래프]
        C2 --> D2[커뮤니티 요약]
        D2 --> E2[그래프 탐색]
    end

Key Differences

Global vs Local Search

graph TD
    A[Query] --> B{검색 타입}
    B -->|Global| C[커뮤니티 요약 탐색]
    B -->|Local| D[엔티티 이웃 탐색]
    C --> E[전체적 관점의 답변]
    D --> F[구체적 사실의 답변]

Experiment Results

Basic experiments were run across 3 commits:

• Indexing takes significantly longer than standard RAG (due to the volume of LLM calls)
• Noticeably better performance on summarization queries compared to standard RAG
• Standard RAG is more efficient for straightforward factual questions

The insights from GraphRAG were later referenced when considering graph-based retrieval options in PlateeRAG/XGen.