Jonathan Larson is a partner data architect at Microsoft Research who leads the team behind GraphRAG, a system that gives language models a structured map of relationships within complex datasets. His research combines graph machine learning, retrieval, evaluation, and agent orchestration to help AI systems understand information dispersed across documents or software repositories.
Before Microsoft, Larson was chief scientist and technical fellow at Sotera Defense Solutions, where he worked on assignment to DARPA. His earlier work also included large-scale data science at Google, Zillow, and the U.S. Army. At Microsoft, he applied network analysis to cybersecurity, search, organizational science, and products including Bing, Viva, and Power BI. His research on collaboration networks examined informal working relationships that organizational charts fail to capture.
That experience informs GraphRAG, which Larson and collaborators introduced in their foundational research paper. The system extracts entities and relationships from unstructured material, organizes them into communities, and uses that structure to answer questions spanning entire datasets. Larson has characterized the project’s ambition as unlocking useful knowledge from complex private information.
- Structured memory for language models. GraphRAG distinguishes local questions answerable from individual passages from global questions requiring connections across an entire corpus. Its graph-based representations make those broader relationships available during retrieval.
- Repository-level code understanding. In an AI Engineer demonstration, Larson showed graph-guided retrieval identifying how movement, obstacles, keyboard input, and rendering interact in a multifile game. The same approach translated the application from Python into working Rust, generated cross-file documentation for Doom, and helped a GitHub Copilot coding agent add jumping through coordinated changes across the codebase.
- LazyGraphRAG. Larson coauthored an approach to reducing GraphRAG’s upfront processing costs by postponing language-model-intensive work until query time. His evaluations compared graph-guided retrieval with vector-based systems across different context-window sizes, challenging the assumption that supplying more text necessarily produces stronger dataset-wide understanding.
- BenchmarkQED and VeriTrail. Larson helped create BenchmarkQED, which generates local and global test questions, evaluates responses, and summarizes or samples datasets. His VeriTrail research, coauthored with Dasha Metropolitansky, addresses hallucination detection and traceability by examining whether generated claims remain grounded in their source material.
