Leveraging Knowledge Graphs and LLMs to Document Large Scale Codebases
summary
Lockheed Martin builds a pipeline that parses multi-language repos with Tree-sitter, extracts import and class-interaction edges into a knowledge graph, ranks nodes by out-degree to find entry points, and feeds concise AST snippets to an LLM to generate UML activity/sequence/class diagrams. It handles codebases from thousands to millions of lines and demonstrates on Spring JDBC (over 128K token worth of files), producing grounded architecture diagrams automatically and reducing manual documentation effort.
AI-written summary · about
chapters (13)
- 0:00Introduction and speaker background
- 0:46Problem with codebase documentation
- 2:36Introducing knowledge graphs for LLMs
- 3:38Parsing code with Tree-sitter
- 4:40Relationships in knowledge graphs
- 6:27Abstract syntax trees and key artifacts
- 7:30Identifying important nodes in graphs
- 9:17Generating UML diagrams
- 11:01Demo with Spring JDBC library
- 12:02Activity diagram walkthrough
- 12:45Sequence diagram walkthrough
- 13:48Class diagram walkthrough
- 14:33Team acknowledgments and conclusion
related talks
30mBuilding Knowledge Graphs With LLMs
20mPractical GraphRAG: Making LLMs smarter with Knowledge Graphs
44mGive Your LLMs a Left Brain
23mKnowledge Graphs and LLMs in Action
52mGive Your LLMs a Left Brain
28mThe Present and Future of AI-Enabled Knowledge Graph Generation
28mGrounded Intelligence Boosting LLMs with Real-Time Graph Context
41mEnhancing LLMs with Graph Technology