microsoft/graphrag

A modular, graph-based Retrieval-Augmented Generation pipeline that uses LLMs to extract knowledge graphs from unstructured text and reason over private data.

What it is

GraphRAG is a data pipeline and transformation suite that extracts structured knowledge-graph data from unstructured text using LLMs, then uses that graph to improve retrieval-augmented answers. It is aimed at teams enhancing an LLM's ability to reason over private or narrative data where flat vector retrieval misses cross-document structure. The README explicitly frames the code as a demonstration of a methodology, not an officially supported Microsoft offering.

Key features

An indexing pipeline that turns unstructured text into a knowledge graph of entities, relationships, and community structures.
Multiple query modes shown in the example notebooks, including global search, local search, DRIFT search, and dynamic community selection.
A command-line workflow via graphrag init, index, update, query, and prompt-tune.
Prompt tuning to adapt extraction prompts to your own data, which the README strongly recommends over out-of-the-box use.
Migration notebooks to move indexes across major versions without full re-indexing.
A modular workspace split into packages for chunking, storage, cache, vectors, input, and LLM access.

Tech stack

Python, constrained to >=3.11,<3.14, organized as a uv workspace monorepo under packages/*.
Vector storage via LanceDB, visible in the example inputs, with Parquet files as pipeline artifacts.
Task automation through poethepoet, versioning through semversioner, and documentation through mkdocs-material.
Developer tooling includes ruff, pyright, and pytest (with asyncio and xdist).
Distributed on PyPI as graphrag; licensed MIT.

When to reach for it

Question answering over a private corpus where global, cross-document questions matter more than single-passage lookups.
Building an explainable knowledge-graph memory structure to augment LLM outputs.
Narrative or complex datasets where similarity-only vector RAG misses thematic structure.

When not to reach for it

Cost-sensitive or small tasks, since the README warns that indexing can be an expensive operation and recommends starting small.
Projects needing an officially supported, production-backed product, which the README explicitly says this is not.
Simple similarity lookups where conventional vector RAG is sufficient and cheaper.
Workflows needing stable configuration across upgrades, since minor version bumps may require re-running init and major bumps may require migration notebooks.

Maturity signal

With over 34,000 stars, an MIT license, creation in March 2024, a last push in July 2026, and a long semversioner history running through the 3.x series, GraphRAG is very actively developed. That said, Microsoft frames it as a research demonstration rather than a supported offering, and warns about indexing cost and breaking changes between versions, so it is best treated as active but research-grade.

Alternatives

LlamaIndex — use it when you want general-purpose document indexing and RAG without committing to graph extraction.
LangChain — use it when you need a broad orchestration framework rather than a focused graph-RAG pipeline.
Neo4j with graph/vector retrieval — use it when you already run a graph database and want to manage the knowledge graph directly there.

Notes

The README carries an explicit warning that indexing can be expensive; reading the documentation and starting small is advised before running at scale.
The project is presented as a methodology demonstration rather than a supported product, with a linked Microsoft Research blog post and an arXiv paper (2404.16130).
Upgrading is not seamless: re-initialize configuration between minor versions and use migration notebooks between major versions.

Wiki: microsoft/graphrag