Dynamic Summaries and Efficient Updates in Knowledge Graphs Integrated with LLMs

Abstract

Low-Code software development hides technical complexity to give users a method for developing software without a prerequisite need for extensive technical knowledge. These methods are powerful, but suffer from the same issues as traditional software development methods with regards to inadequate documentation. CLAIR has since emerged to address these issues through of an LLM-based multi-agent system that automatically generates documentation for Mendix applications. Currently, CLAIR must regenerate all documentation summaries from scratch when a new application version is deployed. This thesis presents CLAIR+, an extension of the original architecture that dynamically updates its hierarchical knowledge graph across versions. The approach is based on the identification of changes between versions, mapping these changes to affected nodes in the knowledge graph and regenerating summaries for the changed components while keeping the rest of the graph in its original state. The goal is to improve scalability and efficiency of the system, without compromising on the quality of generated documentation. The performance of the proposed architecture is validated against the current implementation based on an extensive list of structural and semantic metrics, supplemented by assessments of summaries from experienced Mendix developers.