GraphRAG

GraphRAG is an open-source Python indexing and retrieval pipeline for developers building question answering over their own document collections. It can connect to local models through Ollama proxies, while OpenAI models are its default supported choice. With a remote model provider, text sent for model processing leaves your machine.

Its knowledge graph links entities and relationships extracted from documents, with references back to source passages. Summaries of connected groups supply context for questions about themes across a collection, including questions that require combining information scattered across documents.

Global Search uses those summaries for broad questions. Local Search focuses on an entity and its connections; DRIFT Search adds group context to that exploration. Here, "Local Search" describes retrieval scope. Basic Search uses conventional vector similarity to find relevant passages.

Model compatibility matters. GraphRAG uses LiteLLM for model calls and requires reliable structured JSON responses. Ollama proxy connections can produce malformed responses, so local model suitability depends on meeting those requirements. Prompts can be tuned to the dataset, and a fast indexing method replaces some LLM-based extraction with NLP. LLM indexing can be expensive.

GraphRAG uses the MIT license. It's a Microsoft Research demonstration without official product support and is largely in maintenance mode. Maintainers don't accept new pull requests or develop new features, but perform bug fixes and dependency updates as appropriate, particularly for security vulnerabilities.