Energy consumption forecasting increasingly relies on complex machine learning models such as Genetic Programming‑based symbolic regressors. These models produce accurate predictions but are hard for facility managers and building operators to interpret. Conventional Explainable AI (XAI) dashboards offer feature importance charts, local explanations, etc., yet they demand substantial technical expertise and provide a static interaction model that limits context‑aware queries.\ \ Conversational XAI systems have been proposed to improve usability. Early attempts like TalkToModel used a rigid custom grammar, achieving only 76.8% intent‑parsing accuracy and struggling to generalize across tasks.\ \ This paper introduces the Explainability Assistant, an open‑source conversational XAI framework that leverages function‑calling capabilities of modern Large Language Models (LLMs) to overcome these drawbacks. The architecture consists of:\
- Intent parsing layer: User utterances are sent to an LLM with a predefined function schema; the LLM returns a function call that encodes the user’s intent. This raises intent‑parsing accuracy to 94%.\
- Explanation backend: A collection of callable functions wrapping standard XAI techniques (SHAP, LIME, symbolic regression tree visualizations, etc.).\
- Dialogue manager: Maintains conversational context, allowing users to refine queries step‑by‑step, e.g., “Show the error distribution for the past week” followed by “Which features contribute most to the error?”.\ \ The system requires no task‑specific fine‑tuning; it runs on a generic LLM (e.g., GPT‑4) out‑of‑the‑box. We conducted a comparative user study with eight energy‑domain specialists, who performed identical interpretation tasks using the Explainability Assistant and a traditional XAI dashboard. Findings include:\
- Usability scores improved by ~1.3 points on a 5‑point Likert scale.\
- Task accuracy remained above 96%, with no significant degradation.\
- All participants unanimously preferred the conversational interface for routine interpretability work.\ \ In summary, the Explainability Assistant demonstrates that LLM‑driven function calling can deliver high‑precision intent parsing and flexible natural‑language interaction, substantially lowering the barrier for interpreting energy consumption models.\ \ Review: This work showcases a practical integration of LLMs with classic XAI methods, highlighting the potential to make model explanations more accessible and context‑aware, which is especially valuable for non‑technical stakeholders.