graph TB;
A[Source Documents]
B[Text Chunks]
C[Element Instances]
D[Element Summaries]
E[Graph Communities]
F[Community Summaries]
G[Community Answers]
H[Global Answer]
I[Query]

A -->|extract/chunk text| B
B -->|domain-tailored summarization| C
C -->|domain-tailored summarization| D
D -->|community detection| E
E -->|domain-tailored summarization| F
F -->|query-focused summarization| G
G -->|query-focused summarization| H
I --> G

However, GraphRAG still have certain drawbacks:

  • High cost.

    LLM calls occurs for each community summarization. During query processing, each community report will be processed by LLM again.

  • Slow performance.

    Both community report generation and traversal costs a lot of time.

  • Poor scalability.

    Merging new documents requires reconstruction, limiting the scalability.

RAG