Microsoft GraphRAG is an open-source, graph-augmented retrieval-augmented generation framework developed by Microsoft Research. Unlike naive RAG systems that rely solely on top-k vector similarity over raw document chunks, GraphRAG extracts entities, relationships, and claims from unstructured text to build an interconnected knowledge graph. The system uses the hierarchical Leiden community detection algorithm to group related nodes into semantic clusters across multiple levels of granularity, synthesizing automated community summaries that capture macro-level themes across the entire dataset.
At query time, GraphRAG provides specialized retrieval modes tailored to question complexity: Global Search executes a map-reduce summarization pipeline over precomputed community reports to answer holistic, dataset-wide queries such as identifying overarching trends or systemic risks; Local Search fans out from specific entities into immediate neighbors and relevant text chunks for grounded fact retrieval; and DRIFT search dynamically combines graph traversal with targeted expansion. The pipeline integrates with diverse embedding models, vector stores, and structured storage backends including Parquet and LanceDB.
GraphRAG is best suited for complex enterprise corpora, research archives, intelligence analysis, and multi-document reasoning where standard semantic search misses holistic connections across distributed documents. Deploying GraphRAG requires upfront LLM token investment during graph indexing and community summarization, making prompt caching and structured chunking important operational considerations. Teams evaluating graph-native context infrastructure can compare GraphRAG with hybrid graph-memory engines like Cognee, LightRAG, and Zep based on indexing throughput, incremental update support, and real-time query latency requirements.
