GraphRAG Workflow

GraphRAG extracts an entity graph from unstructured text using LLM prompting, clusters it into communities, and routes each query to local (entity-level) or global (corpus-level) retrieval.

%%{init: {'theme':'base','themeVariables':{'fontFamily':'Helvetica, Arial, sans-serif','fontSize':'14px','primaryColor':'#E8EEF7','primaryBorderColor':'#00796B','primaryTextColor':'#004D40','lineColor':'#00796B','clusterBkg':'#F1F8E9','clusterBorder':'#AED581'}}}%%
flowchart TB
  classDef input fill:#E0F2F1,stroke:#00695C,stroke-width:1.5px,color:#004D40
  classDef process fill:#E8F5E9,stroke:#2E7D32,stroke-width:1.5px,color:#1B5E20
  classDef store fill:#FFF3E0,stroke:#E65100,stroke-width:1.5px,color:#BF360C
  classDef output fill:#E3F2FD,stroke:#1565C0,stroke-width:1.5px,color:#0D47A1

  subgraph Index[Build time]
    direction LR
    Text["Domain text<br/>(Markdown chapters)"]:::input --> Extract["LLM entity &<br/>relation extraction"]:::process
    Extract --> Graph[("Extracted<br/>entity graph")]:::store
    Graph --> Community["Community<br/>detection"]:::process
    Community --> Reports["Community<br/>summaries"]:::store
  end

  subgraph Query[Query&nbsp;time]
    direction LR
    Q(["User query"]):::input --> Route["Local vs. global<br/>route"]:::process
    Route --> Ctx["≈ 2,000–8,000 tokens<br/>of retrieved context"]:::process
    Ctx --> LLM["LLM<br/>(Claude)"]:::process
    LLM --> Ans(["Answer"]):::output
  end

  Graph -.-> Route
  Reports -.-> Route

Typical retrieved context: ~2,000–8,000 tokens. Build cost: LLM passes over the entire corpus to extract entities, relations, and community summaries. Strength: produces corpus-level answers that pure RAG cannot. Weakness: dynamic extraction reintroduces structure that was already implicit in the source — then asks the retrieval LLM to use it. Extracted edges can disagree with ground-truth structure (measurable hallucination rate).