Microsoft GraphRAG

Overview

GraphRAG is a modular graph-based Retrieval-Augmented Generation (RAG) system developed by Microsoft Research. It extracts meaningful, structured data from unstructured text using LLMs to create a knowledge graph, then uses these connections to answer questions that span many documents or require thematic understanding. Unlike traditional vector-based RAG, GraphRAG excels at answering abstract questions like "What are the top themes in this dataset?" by leveraging community detection and hierarchical summarization.

Problem Addressed

Key Statistics

Key Features

Indexing Pipeline

• Entity extraction: LLM-powered extraction of entities, relationships, and claims from raw text
• Community detection: Graph-based clustering to identify related entity communities
• Hierarchical summarization: Community reports generated at multiple levels of granularity
• Chunk embedding: Text chunks embedded for vector similarity search
• Entity embedding: Entities embedded for semantic retrieval
• LLM caching: Cached completions for idempotent, resilient indexing
• Configurable workflows: Modular pipeline with customizable steps and prompts

Query Mechanisms

• Local Search: Combines AI-extracted knowledge graph with text chunks for entity-specific questions
• Global Search: Map-reduce over community reports for dataset-wide thematic questions
• DRIFT Search: Expands local search breadth using community insights for comprehensive answers
• Basic Search: Vector RAG baseline for comparison (top-k chunk retrieval)
• Question Generation: Generates follow-up questions for deeper investigation

Architecture

• Monorepo structure: Modular packages (graphrag-cache, graphrag-chunking, graphrag-common, graphrag-input, graphrag-llm, graphrag-storage, graphrag-vectors)
• Factory pattern: Extensible providers for models, storage, cache, vectors, input readers
• Parquet outputs: Indexes stored as Parquet tables for efficient querying
• CLI and Python API: Multiple interfaces for indexing and querying

Model Support

• OpenAI: Direct API integration
• Azure OpenAI: Full support with managed identity authentication
• LiteLLM wrapper: Extensible to additional providers
• Prompt tuning: Guidance for domain-specific prompt optimization

Storage Backends

• File storage: Local filesystem for development
• Azure Blob Storage: Cloud storage integration
• CosmosDB: Distributed database support
• Vector stores: LanceDB, Azure AI Search, CosmosDB vector search

Technical Architecture

Pipeline Flow

Knowledge Model

Factory Extensions

Installation and Usage

Installation

# Using pip
pip install graphrag

# Using uv
uv pip install graphrag

Quick Start

# Initialize workspace
mkdir graphrag_project && cd graphrag_project
graphrag init

# Configure API key in .env
# GRAPHRAG_API_KEY=<your-api-key>

# Add documents to ./input directory
curl https://www.gutenberg.org/cache/epub/24022/pg24022.txt -o ./input/book.txt

# Run indexing
graphrag index

# Query with Global Search (thematic questions)
graphrag query "What are the top themes in this story?"

# Query with Local Search (entity-specific questions)
graphrag query "Who is Scrooge and what are his main relationships?" --method local

Python API

from graphrag.api import build_index, local_search, global_search

# Index documents
await build_index(root="./graphrag_project")

# Local search for entity questions
result = await local_search(
    root="./graphrag_project",
    query="What are the healing properties of chamomile?"
)

# Global search for thematic questions
result = await global_search(
    root="./graphrag_project",
    query="What are the main themes across all documents?"
)

Azure OpenAI Configuration

# settings.yaml
models:
  default_chat_model:
    type: chat
    model_provider: azure
    model: gpt-4.1
    deployment_name: <AZURE_DEPLOYMENT_NAME>
    api_base: https://<instance>.openai.azure.com
    api_version: 2024-02-15-preview
    auth_type: azure_managed_identity  # Optional: for managed auth

Relevance to Claude Code Development

Direct Applications

1. Knowledge-Augmented Skills: GraphRAG patterns could inform how Claude Code skills organize and retrieve reference documentation, using graph structure instead of flat files.

2. Cross-Document Reasoning: The community detection and hierarchical summarization approach provides a model for answering questions that span multiple skill reference files.

3. Thematic Query Support: Global search mechanism could inspire how Claude Code answers abstract questions about a codebase ("What patterns does this project follow?").

4. Context Window Optimization: Community reports provide compressed summaries that could reduce token usage while maintaining comprehensive coverage.

5. Caching Patterns: LLM caching strategy applicable to Claude Code for expensive operations.

Patterns Worth Adopting

1. Multiple Search Strategies: Local vs Global vs DRIFT search demonstrates value of query-type-specific retrieval strategies.

2. Community Detection: Graph clustering to identify related concepts could improve skill organization and cross-referencing.

3. Hierarchical Summarization: Multi-level community reports enable both detailed and overview queries on same index.

4. Factory Pattern: Extensible providers for all subsystems enable customization without core changes.

5. Parquet Outputs: Columnar format for index storage enables efficient analytical queries.

Integration Opportunities

1. MCP Server: GraphRAG could be exposed as an MCP tool for Claude Code to query indexed knowledge bases.

2. Skill Indexing: Index skill reference documentation with GraphRAG for enhanced retrieval in complex multi-skill queries.

3. Codebase Understanding: Index codebase documentation/comments to answer architectural questions.

4. Research Aggregation: Index research entries (like this one) for cross-resource thematic queries.

Comparison: GraphRAG vs Traditional RAG

Cost Considerations

The README explicitly warns: "GraphRAG indexing can be an expensive operation." Best practices:

• Start with small test datasets
• Use inexpensive/fast models for initial experimentation
• Leverage LLM caching to avoid redundant API calls
• Consider update strategies before large indexing jobs

References

Research Method: Information gathered from official GitHub repository README, RAI transparency document, documentation pages (query overview, index overview, architecture), GitHub API for statistics, PyPI API for package info and download statistics. All claims verified against primary sources.