GraphRAG is a new approach developed by Microsoft Research to improve the performance of Retrieval-Augmented Generation (RAG) models, particularly when addressing complex queries that require an understanding of a large body of documents. Traditional RAG methods retrieve data points individually, which falls short for broad questions like identifying patterns across thousands of reports. GraphRAG overcomes this by creating a hierarchical knowledge graph before any query is posed. This graph organizes information into a network of entities and relationships, making it easier to navigate and synthesize insights from extensive data. The process begins with segmenting source documents into manageable chunks of around 600 tokens, ensuring semantic continuity through overlapping segments. Each chunk is then analyzed by a language model to extract entities, relationships, and claims. The extraction process uses multipartite prompts and a "gleaning" technique to minimize the risk of missing important information. Claims are anchored to the source text, allowing for verification and reducing inaccuracies. Once the entities and relationships are extracted, GraphRAG consolidates duplicate mentions into a single node using entity resolution algorithms. This creates a multigraph that represents interconnected data points. The next step involves community detection using the Leiden algorithm, which organizes the graph into clusters based on the density of connections. This hierarchical structure allows for efficient navigation and summarization of topics. The algorithm's ability to adjust granularity through a resolution parameter adds flexibility in tailoring the level of detail based on specific queries.