← back
GraphRAG methods to create optimized LLM context windows for Retrieval — Jonathan Larson, Microsoft
Takeaway
Graph-structured memory turns RAG from snippet retrieval into repository-scale reasoning that survives multi-file edits like adding jump mechanics to Doom.
Summary
- Jonathan Larson (Microsoft Research) demos GraphRAG for code: structures repository understanding enabling both local (per-file) and global (cross-module) queries.
- On a 200-line Python sidescroller, vanilla RAG produces vague 'it's a game' answers while GraphRAG-for-code identifies player jumping, obstacle scrolling, spacebar control—correctly translates the codebase to compilable Rust.
- Scaled to 100k-line Doom codebase: generated module-level docs and added a jumping feature requiring multi-file edits that plain LLMs break.
- Announces open-source release of benchmark QED and previews LazyGraph evolution with new benchmark numbers; key thesis: structured LLM memory + agents > flat embeddings.
graphragcode-understandingmicrosoft
Original description
Jonathan Larson is a Senior Principal Data Architect at Microsoft Research working in Special Projects. He currently leads a research team focused on the intersection of graph machine learning, LLM memory representations, and LLM orchestration. His research has led to shipping new features in Bing, Viva, PowerBI. He also shipped new tools to combat tech fraud. Many of the supporting libraries have been open sourced in collaboration on GitHub. Prior to joining Microsoft, Jonathan was Chief Scientist and Technical Fellow at Sotera Defense Solutions on assignment to DARPA, and led a variety of research across several programs. Jonathan has also led large-scale data science efforts at Google, Zillow, and the US Army. Early in his career, he also worked several startups and incubators. About Jonathan Larson Jonathan Larson is a Senior Principal Data Architect at Microsoft Research working in Special Projects. He currently leads a research team focused on the intersection of graph machine learning, LLM memory representations, and LLM orchestration. Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter