GraphRAG Documentation (Microsoft Research)
by Microsoft Research
Build a knowledge graph over your corpus so RAG can answer whole-corpus questions.
Overview
Microsoft Research's GraphRAG documentation covers a modular, graph-based alternative to vector-only retrieval: instead of fetching chunks by embedding similarity, GraphRAG first uses an LLM to build a knowledge graph over your corpus, then answers from that structure. The Indexing section walks the four-stage pipeline — slicing source documents into TextUnits so references stay fine-grained, extracting entities, relationships and claims from each unit, hierarchically clustering those entities into communities with the Leiden technique, and generating bottom-up summaries of each community and its constituents — alongside pages on architecture, dataflow, indexing methods, input and output formats, and ingesting custom graphs. The Query section documents four retrieval modes and when each applies: Global Search for holistic questions about the corpus as a whole, which chunk-and-embed RAG structurally cannot answer; Local Search for reasoning about a specific entity by fanning out to its neighbours; DRIFT Search, a local approach enriched with community context; and Basic Search for plain vector similarity. Prompt Tuning covers both automatic and manual adaptation of the extraction prompts, which the maintainers say is necessary for good results on your own data, and Configuration covers the init command, detailed settings and language-model selection, followed by a CLI reference and notebooks. The project is MIT-licensed with roughly 35.2k GitHub stars, shipped v3.1.1 on 18 July 2026, and is grounded in the arXiv paper 2404.16130.
At a Glance
- Topic
- RAG
- Level
- Advanced
- Format
- Documentation
- Cost
- Free
- Duration
- ~4-6 hours for the quickstart plus a first real indexing run
- Provider
- Microsoft Research
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Run the CLI quickstart: graphrag init, index a corpus, then query it
- ✓Understand the four-stage indexing pipeline from TextUnits to community summaries
- ✓Choose between Global, Local, DRIFT and Basic search for a given question
- ✓Tune the entity and claim extraction prompts automatically or by hand for your domain
- ✓Read the parquet outputs: entities, relationships, communities, claims and text units
- ✓Budget an indexing run, since graph extraction calls the LLM per text unit
- ✓Configure the language model and vector store through settings.yaml and the init command
Highlights
- •Answers corpus-level questions such as 'what are the main themes here?' that chunk-and-embed RAG structurally cannot
- •The Leiden-clustered community summaries are the actual novelty, and they are documented end to end rather than hand-waved
- •MIT-licensed and still actively released — v3.1.1 shipped 18 July 2026, with breaking-change notes for version bumps
- •Unusually honest about cost: the docs tell you to start on the tutorial dataset with a cheap model before committing to a big index
- •Practitioners on Hacker News find the Microsoft implementation hard to read and hack — nano-graphrag exists as an ~800-line readable alternative, which is worth knowing before you fork it
Who It’s For
Best For
- ✓RAG engineers whose users ask summarising questions across an entire document set
- ✓Teams evaluating graph RAG before committing to building one in-house
- ✓Analysts working over private corpora such as reports, transcripts or incident logs
- ✓Anyone who has hit the ceiling of top-k chunk retrieval and needs structure
Prerequisites
- •Python and virtual environments; the tool is installed and driven from the command line
- •A working RAG mental model — chunking, embeddings, top-k retrieval — since this is the contrast being drawn
- •An LLM API budget, because indexing is token-heavy and the docs warn about it explicitly
FAQ
What is GraphRAG Documentation (Microsoft Research)?
Microsoft Research's official documentation for GraphRAG, a modular graph-based RAG system that builds a knowledge graph and hierarchical community summaries over your corpus instead of retrieving isolated chunks. Aimed at RAG engineers whose users ask corpus-wide questions that top-k vector retrieval answers badly. After working through it you can index a private document set from the CLI, choose between global, local, DRIFT and basic search, tune the extraction prompts to your own domain, and estimate what an indexing run will actually cost.
Is GraphRAG Documentation (Microsoft Research) free?
GraphRAG Documentation (Microsoft Research) is free to access.
What level is GraphRAG Documentation (Microsoft Research) for?
GraphRAG Documentation (Microsoft Research) is aimed at a advanced audience. Recommended background: Python and virtual environments; the tool is installed and driven from the command line, A working RAG mental model — chunking, embeddings, top-k retrieval — since this is the contrast being drawn, An LLM API budget, because indexing is token-heavy and the docs warn about it explicitly.
How long does GraphRAG Documentation (Microsoft Research) take?
Expect roughly ~4-6 hours for the quickstart plus a first real indexing run. Most learners work through it at their own pace.
What will I learn from GraphRAG Documentation (Microsoft Research)?
You'll learn: Run the CLI quickstart: graphrag init, index a corpus, then query it; Understand the four-stage indexing pipeline from TextUnits to community summaries; Choose between Global, Local, DRIFT and Basic search for a given question; Tune the entity and claim extraction prompts automatically or by hand for your domain; Read the parquet outputs: entities, relationships, communities, claims and text units; Budget an indexing run, since graph extraction calls the LLM per text unit; Configure the language model and vector store through settings.yaml and the init command.
Topics
Sources
This page was written from 4 sources, 3 on domains other than microsoft.github.io.