An Interactive LLM-based Conversational Agent for Complex Data Analysis

Abstract

Complex multivariate datasets—characterized by complex parent-child structures and rich attributes such as hierarchies and networks—pose challenges for intuitive exploration and analysis. This work presents an interactive visualization system integrated with a conversational agent (chatbot) to support natural language interaction with such data, especially for domain experts. Users can upload datasets, issue natural language queries, manipulate interface elements (e.g., buttons, panels), and generate custom visualizations including force-directed graphs, circle-packing layouts, and tabular charts. These features enhance data interpretability and engagement. The system includes a robust NLP pipeline based on DistilBERT for intent classification, optimized through data balancing and retraining. Visualizations, rendered in real time with Plotly and D3.js in a Dash interface, support interactions such as zooming, panning, node selection, and dynamic color mapping via language commands. A Retrieval-Augmented Generation (RAG) pipeline enriches chatbot responses using contextual information from uploaded documents. The system also supports misclassification reporting to iteratively refine the NLP model. It handles large-scale hierarchical data efficiently and has been validated on examples like organizational charts and threaded discussions. Notable features include real-time GUI customization, multi-turn conversational support, and popup visualizations from selected data subsets using intuitive queries (e.g., “Show toxicity distribution”). User testing showed high satisfaction among experts, while novices noted a steeper learning curve during onboarding.