GraphRAG — Knowledge Graph RAG by Microsoft

What it is

GraphRAG is an open-source tool by Microsoft Research that builds knowledge graphs from unstructured documents for smarter retrieval-augmented generation. Unlike traditional RAG that retrieves text chunks, GraphRAG extracts entities and relationships, organizing them into a graph structure that supports both entity-focused local search and holistic global search.

GraphRAG is designed for researchers and developers building RAG systems that need to answer questions requiring synthesis across multiple documents, not just finding a single relevant passage.

How it saves time or tokens

Traditional vector-based RAG struggles with questions that span multiple documents or require understanding relationships between entities. GraphRAG solves this by pre-indexing documents into a knowledge graph during the indexing phase. At query time, the graph structure enables answering questions like 'what are the key themes across all documents' without scanning every chunk. The estimated token cost for this workflow is around 1,783 tokens.

How to use

1. Install GraphRAG:

pip install graphrag

1. Initialize and index your documents:

graphrag init --root ./my-project
# Place documents in ./my-project/input/
graphrag index --root ./my-project

1. Query with local or global search:

# Local search (entity-focused)
graphrag query --root ./my-project --method local \
  'What are the main findings about transformer architectures?'

# Global search (holistic summary)
graphrag query --root ./my-project --method global \
  'What are the key themes across all documents?'

Example

# Python API usage
from graphrag.query.cli import run_local_search, run_global_search

# After indexing, query programmatically
result = run_local_search(
    root_dir='./my-project',
    query='Who are the key researchers in attention mechanisms?'
)
print(result.response)

# Global search for cross-document synthesis
result = run_global_search(
    root_dir='./my-project',
    query='Summarize the evolution of language models'
)
print(result.response)

Related on TokRepo

• RAG tools — retrieval-augmented generation frameworks
• Knowledge graph tools — entity and relationship extraction

Common pitfalls

• Not configuring the LLM provider in settings.yaml before running the indexing step
• Running indexing on very large document sets without estimating the LLM API costs first
• Using local search when the question requires cross-document synthesis (use global search instead)