Revamping Software Documentation: A New Approach to UML Diagram Generation

Automating software documentation is no small feat, especially large codebases where context limits of Large Language Models (LLMs) often fall short. A new agentic architecture, however, might just provide the breakthrough needed. This innovative system efficiently generates UML diagrams from source code, addressing scalability issues head-on.

Agentic Architecture in Action

The architecture introduces a hierarchy of five specialized agents: PlannerAgent, AnalyzerAgent, DiagramAgent, CorrectorAgent, and DependencyAnalyzerAgent. Built using the Claude Agent SDK, each agent handles a specific cognitive subtask, paving the way for precise and efficient documentation.

What sets this system apart is its deterministic, importance-weighted Intermediate Representation (IR) compaction layer. This layer compacts full project IRs into diagram-specific views that fit within token constraints, all without requiring additional LLM calls. The process is swift, completing in milliseconds, which is a significant advancement in current methodologies.

Performance Metrics and Results

The system was rigorously tested across 12 open-source repositories spanning four programming languages: Java, JavaScript, PHP, and Python. It generated seven types of UML diagrams, resulting in 84 observations assessed using five automated metrics.

The results speak volumes. The system demonstrated high syntactic validity, with a mean score of 91.5%, and both component and deployment diagrams reached an impressive 100%. Relationship precision was strong at 0.858, and structural quality scores were consistently high, averaging 81.7 out of 100 across languages. Although the entity recall averaged only 0.313, this reflects a deliberate choice to prioritize architectural integrity over exhaustive coverage.

Why This Matters

In a world where software documentation is essential yet often neglected, this approach offers a compelling solution. By automating the generation of UML diagrams with such precision and efficiency, it could revolutionize how developers approach documentation.

But here's the real question: Could this architecture redefine the standards for documentation tools? With the results showing stability regardless of the scale of IR entities, ranging from 31 to 4,578, it's clear that scalability concerns are being addressed effectively. This innovation not only enhances documentation practices but also empowers developers to focus on what truly matters: coding.