This paper introduces a Graph RAG approach for query-focused summarization (QFS) over private text corpora. The method combines the strengths of retrieval-augmented generation (RAG) and QFS to address the limitations of both approaches. The Graph RAG approach builds a graph-based text index using an LLM, first deriving an entity knowledge graph from source documents and then pre-generating community summaries for closely-related entities. Given a question, each community summary is used to generate a partial response, which are then summarized into a final response. This approach is particularly effective for global questions over large datasets, as it leverages the modularity of graphs and community detection to partition the data into manageable parts. The method is evaluated on two datasets, each containing around a million tokens, and shows significant improvements in comprehensiveness and diversity compared to a naive RAG baseline. The approach also outperforms text summarization methods in terms of token efficiency. The Graph RAG approach is implemented in Python and is available as an open-source project. The paper also discusses related work, including other RAG approaches and graph-based methods, and highlights the potential of the Graph RAG approach for future research and applications in data sensemaking.This paper introduces a Graph RAG approach for query-focused summarization (QFS) over private text corpora. The method combines the strengths of retrieval-augmented generation (RAG) and QFS to address the limitations of both approaches. The Graph RAG approach builds a graph-based text index using an LLM, first deriving an entity knowledge graph from source documents and then pre-generating community summaries for closely-related entities. Given a question, each community summary is used to generate a partial response, which are then summarized into a final response. This approach is particularly effective for global questions over large datasets, as it leverages the modularity of graphs and community detection to partition the data into manageable parts. The method is evaluated on two datasets, each containing around a million tokens, and shows significant improvements in comprehensiveness and diversity compared to a naive RAG baseline. The approach also outperforms text summarization methods in terms of token efficiency. The Graph RAG approach is implemented in Python and is available as an open-source project. The paper also discusses related work, including other RAG approaches and graph-based methods, and highlights the potential of the Graph RAG approach for future research and applications in data sensemaking.