microsoft/graphrag

GraphRAG is a Python framework from Microsoft Research that uses knowledge graphs to improve retrieval-augmented generation (RAG) systems. Traditional RAG pipelines retrieve relevant text chunks from a document corpus based on similarity to a query and pass them to a language model. GraphRAG extends this by first building a knowledge graph from the source documents, extracting entities, relationships, and hierarchical community summaries, and then using that structured graph to retrieve information at query time. The key advantage is in handling questions that require understanding how different pieces of information relate to each other across a large corpus, rather than just finding the single most similar passage. Because the graph captures entity relationships and organizes content into communities at multiple levels of granularity, it can answer global questions about the dataset that simple vector search would struggle with. The framework supports two query modes: local search, which finds precise answers about specific entities, and global search, which reasons across the full dataset using the hierarchical community summaries built during indexing. The indexing pipeline processes documents through an extraction stage that identifies entities and relationships using a language model, then organizes these into a community hierarchy. At query time, the system uses this pre-built structure rather than raw text alone. When to use it: GraphRAG is suited for organizations that need to query large private document collections where questions often span multiple documents or require understanding how topics connect. It is most powerful when the corpus has complex relational structure, legal documents, research papers, internal knowledge bases, and simple keyword or embedding search produces incomplete answers. The stack is Python, with integration into Azure OpenAI and other LLM providers.