GraphRAG

GraphRAG is a modular graph-based retrieval-augmented generation system from Microsoft Research, with 36,000+ stars on GitHub, designed to answer questions that ordinary RAG cannot over private data.

Overview

Standard RAG retrieves a handful of similar chunks, which fails on global questions like “what are the main themes across these thousand documents?” GraphRAG takes a structural approach: it uses an LLM to extract entities and relationships from your corpus, builds a knowledge graph, detects communities within it and pre-generates summaries of each community. At query time, global search reasons over those community summaries for corpus-wide sense-making, while local search drills into specific entities and their neighborhoods. The result is a pipeline and transformation suite aimed at narrative, private datasets — investigative documents, research collections, internal reports. Microsoft notes the project is now largely in maintenance mode, receiving bug fixes rather than new features, and warns that indexing is an LLM-driven operation that can be expensive, so start small. The repository remains available and was updated through October 2026.

Key Features

• Graph-based indexing – LLM extraction of entities and relationships builds a knowledge graph from raw text.
• Community summaries – hierarchical community detection with generated summaries powers global questions.
• Global and local search – corpus-wide sense-making plus entity-focused drill-down in one system.
• Built for private data – designed for narrative corpora you cannot send to a public chatbot.
• Modular pipeline – configurable indexing workflows, prompts and storage backends.
• CLI quickstart – a guided command-line path from documents to a queryable index.
• Research pedigree – from Microsoft Research, with a published paper and detailed documentation.

💳 Pricing

GraphRAG is free and open source under the MIT license. The real cost is inference: indexing and querying consume LLM API calls, and large indexing runs can be expensive, so budget and start with a small corpus.

Who Is It For

Analysts, investigators and research teams asking sense-making questions over large private document collections, where the answer lives in themes across documents rather than in any single passage.