GraphRAG methods to create optimized LLM context windows for Retrieval
summary
Demonstrates that structured LLM memory with graphs actually enables repository-level reasoning: graph for code translates a 200-line Python game into working Rust and modifies the 100k-line Doom codebase to add jumping across many files without breaking builds; introduces open-source Benchmark QED (autoQ, autoE, autoD) for local vs global query evaluation; shows lazy graph outperforms vector RAG (~90%+ wins on data-local/global benchmarks) while costing about one‑tenth of 1M-token context runs and is landing in Azure Local and Microsoft Discovery.
AI-written summary · about
chapters (8)
- 0:00Introduction to GraphRAG
- 1:17GraphRAG for code understanding
- 4:13Code translation with GraphRAG
- 6:03Applying GraphRAG to Doom
- 7:36Feature development with agents
- 9:39BenchmarkQED overview
- 11:55LazyGraphRAG performance
- 13:43Azure integration and summary
related talks
26mPractical GraphRAG Making LLMs smarter with Knowledge Graphs
53mPractical GraphRAG - Making LLMs smarter with Knowledge Graphs
49mPower up LLMs to be Factual Ninjas using GraphRAG
28mEnhancing LLMs with Graph Technology
31mEnhancing RAG-based apps by constructing and leveraging knowledge graphs with open-source LLMs
27mGraphReader Agentic RAG: Rethinking Long-Context Retrieval Systems
06mCloud-Native GraphRAG: AI’s Logical Edge
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance