Microsoft GraphRAG

Microsoft GraphRAG is an open-source retrieval-augmented generation system that builds a knowledge graph from your documents, then uses community detection and pre-computed summaries to answer broad questions traditional RAG cannot.

What It Is

Plain RAG was built for one kind of question: “find me the passage that answers this.” It splits documents into chunks, embeds them, retrieves the closest few, and stitches them into a prompt. That works when the answer lives in a single passage. It fails when the question is broader — “what are the main themes across this dataset?” or “how do these ideas connect across the whole corpus?” — because no single chunk holds the answer. Microsoft GraphRAG was published to handle exactly that class of question.

Instead of treating documents as isolated chunks, GraphRAG runs an LLM over every chunk twice. The first pass extracts entities and relationships, building a knowledge graph where nodes are concepts (people, companies, ideas, events) and edges describe how they relate. The second pass groups densely connected nodes into communities. According to Microsoft GraphRAG concepts, this grouping uses the Leiden algorithm — a hierarchical, recursive method that partitions the graph at multiple resolutions, so you get coarse top-level themes and fine-grained sub-clusters inside the same structure.

Once the graph and communities exist, the system pre-computes a written summary for each community. At query time, two retrieval modes become possible. Local search starts from a specific entity and walks outward through its neighbours — useful for “tell me about X.” Global search reads the community summaries themselves and asks the model to combine them — useful for “what are the patterns across everything?” The graph and summaries are built once during indexing; queries are cheap, but indexing is not. According to Microsoft’s GraphRAG repository, the current release adds an --estimate-cost CLI flag so teams can preview indexing spend before committing.

How It’s Used in Practice

Most teams reach for GraphRAG when they need to query a fixed body of documents that is too large for any single prompt — a year of internal reports, a corpus of regulatory filings, a research literature collection — and where the questions span the whole set rather than asking about one passage. The workflow is consistent: clone the repository, point the indexer at a folder of source files, and let it run. Indexing pulls every chunk through the LLM twice (entities, then summaries), so the upfront bill scales with corpus size and chunk strategy, not with how many people will eventually query the result.

Pro Tip: Run --estimate-cost on a small sample before you index the whole corpus. According to Microsoft’s GraphRAG repository, that flag exists precisely because production indexing is the expensive step — not querying. If the estimate looks painful, drop entity types you do not actually search on. Each unused entity class adds tokens you never recover.

When to Use / When Not

Common Misconception

Myth: Microsoft GraphRAG is a hosted service you can call from an API. Reality: GraphRAG is an MIT-licensed Python library you install and run yourself. According to Microsoft’s GraphRAG repository, there is no managed endpoint — indexing happens on your infrastructure with your LLM credentials, which is why per-corpus cost varies so widely between teams.

One Sentence to Remember

Pick GraphRAG when the question is shaped like “what are the patterns across everything,” and budget for the indexing pass before you fall in love with the demo — the query side is cheap, the build side is where the cost lives.

FAQ

Q: How is Microsoft GraphRAG different from regular RAG? A: Standard RAG retrieves text chunks based on similarity. GraphRAG first extracts a knowledge graph and pre-summarizes connected communities, so it can answer dataset-wide questions where no single chunk holds the answer.

Q: Is Microsoft GraphRAG free to use? A: The code is MIT-licensed and free. Running it is not — every chunk is processed by an LLM during indexing, so the real cost is whatever your model provider charges to extract entities and summarize communities.

Q: Should I use GraphRAG or LazyGraphRAG? A: GraphRAG pays cost upfront during indexing; LazyGraphRAG defers most of that work to query time. According to Microsoft Research, the lazy variant cuts indexing dramatically when you do not need every community pre-summarized.

Sources

• Microsoft’s GraphRAG repository: microsoft/graphrag — A modular graph-based RAG system - canonical implementation, current release notes, and the --estimate-cost CLI flag
• “From Local to Global” arXiv: From Local to Global: A GraphRAG Approach to Query-Focused Summarization - original method paper that introduced the entity-graph + community-summarization pattern