GraphRAG

Microsoft Research built GraphRAG to answer questions that ordinary retrieval augmented generation cannot, the global ones like what are the main themes in this corpus, where no single passage contains the answer. An indexing pipeline uses an LLM to extract entities and relationships from raw text into a knowledge graph, clusters it into hierarchical communities with the Leiden algorithm, and pre-writes summaries for each community; at query time, global search reasons over those summaries while local search fans out from specific entities, and DRIFT search combines the two approaches. The Python package installs from PyPI with a CLI covering init, index, and query, with configuration in YAML. Indexing is LLM-intensive and typically points at OpenAI or Azure OpenAI endpoints, so costs scale with corpus size and offline use is not the default path. Released under MIT with an accompanying research paper, the roughly 35,000-star project effectively defined the graph RAG category in 2024, and it remains the reference implementation that researchers and enterprises benchmark against.