Developing and Evaluating an LLM-based Agent for ExplorViz Using CopilotKit
ExplorViz is an interactive software visualization tool that supports program comprehension by presenting complex software systems using a 3D city metaphor and communication links. While these visualizations provide a rich overview, they also introduce an interaction challenge: users must navigate a three-dimensional landscape, interpret architectural structures, and translate observations into concrete follow-up actions. This thesis designs and implements a large-language-model (LLM) assistant embedded in the ExplorViz frontend that allows users to ask questions about the currently visualized system and trigger visualization actions through natural language. The assistant combines selective, privacy-aware snapshots of the current visualization state with deterministic, tool-based actions for querying landscape structures, highlighting landscape components, camera controls, and settings access. In addition, an experimental workflow supports agent-initiated edits with explicit user confirmation. A controlled usability study with 11 participants evaluated both comprehension support and action execution. Participants rated the assistant’s generated summaries and explanations as largely correct (mean ratings 4.64/5 and 4.55/5) and reported high ease-of-use for deterministic actions such as bulk highlighting and theme creation (mean 4.82/5). In contrast, open-ended agent-driven editing showed mixed expectation match (mean 3.00/5), indicating the need for stronger guardrails and feedback mechanisms. Overall, the assistant was perceived as usable and promising for reducing interaction overhead during exploratory program comprehension tasks.
Preview
Rights
Use and reproduction:
Please note that individual components of the publication may be subject to other licensing or copyright conditions.
