Microsoft GraphRAG

What it is

Microsoft's open-source GraphRAG package. Uses an LLM to extract entities and relationships, builds hierarchical communities, then summarizes each community as a retrievable unit. Excellent for whole-corpus, holistic questions.

How Vaaani uses it

• Whole-corpus summary questions ('what are the main themes in this 500-doc set?')
• Investigations across many interconnected sources
• Multi-hop reasoning where vector RAG fails
• Replacing manual literature reviews

Why it makes the cut

When the customer asks questions that no single document answers, MS GraphRAG genuinely shines. It's the closest thing to 'understanding' a corpus.

Sample code

# config: settings.yaml
llm:
  type: openai_chat
  model: gpt-4o
embeddings:
  type: openai_embedding
  model: text-embedding-3-large
chunks:
  size: 1200