Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models

The Explainability Assistant is an open-source conversational XAI system that leverages LLM function-calling capabilities to allow non-technical users, such as facility managers, to interactively interpret complex machine learning models like Genetic Programming-based symbolic regressors used for energy consumption forecasting.

Unlike traditional static dashboards that require substantial technical expertise, this interface enables flexible, natural language interactions for tasks such as feature importance analysis, counterfactual reasoning, and what-if scenario simulations without the need for task-specific fine-tuning.

Key performance and usability metrics include: 94% intent-parsing accuracy, significantly surpassing the 76.8% accuracy of previous grammar-based systems like TalkToModel. 100% task accuracy among energy domain specialists using the conversational interface, compared to 93% for traditional dashboards. * Unanimous preference from expert evaluators for the conversational interface due to superior usability and ease of use, while maintaining appropriate epistemic caution regarding model outputs.

Explainability Assistant presents a conversational XAI interface for interpreting energy-consumption forecasting models, with a specific focus on models generated by Genetic Programming. Rather than relying on fixed dashboards that display feature-importance plots, residual charts, or model trees, the work frames interpretation as an interactive dialogue: users can ask natural-language questions about model behavior, request explanations in more accessible terms, and iteratively probe why a particular forecast was produced. This is especially relevant for GP-based forecasters, which may learn symbolic or programmatic relationships that are compact and expressive but not immediately legible to non-technical stakeholders such as facility managers, energy operators, or policy users.

The key contribution is the treatment of explainability as an interaction-design problem rather than purely a post-hoc visualization problem. The system bridges the gap between the internal logic of an evolved forecasting model and the practical questions a user actually needs answered: which inputs mattered, whether the model is responding to plausible demand drivers, and how to interpret a prediction in operational terms. By supporting follow-up questioning and context-specific exploration, the interface helps users build a working mental model of the forecaster instead of passively consuming a static explanation.

This matters because energy forecasting increasingly supports demand management, efficiency planning, load shifting, and grid-related decisions, yet the people who must act on those forecasts are often not data scientists. A conversational XAI layer can improve trust, reduce misinterpretation, and make it easier to detect when a model is relying on spurious or unintuitive signals. More broadly, the material offers a useful pattern for applying dialogue-based explainability to other forecasting or optimization systems where transparency is needed for operational, financial, or policy decisions.