Phebe: A Multi-Agent Natural Language Interface for Interactive Data Visualization

Most data visualization tools require technical expertise, creating significant accessibility barriers for non-technical users. We present Phebe, a web-based platform that enables users to upload tabular datasets, perform AI-guided data cleaning, and generate interactive visualizations by expressing analytical needs in plain English. The system is built around a multi-agent architecture powered by large language models (LLMs), comprising an Orchestrator Agent for intent classification, a Text-to-SQL Agent with execution-based self-correction, an Analysis Agent for statistical computations, and a Chart Recommendation Agent for automated visual encoding selection. An end-to-end pipeline integrates CSV ingestion with robust validation, a six-category data quality subsystem, SQL generation against user-uploaded datasets stored in MySQL, and a D3.js-based rendering layer supporting seven chart types, a drag-and-drop dashboard, and PDF export. Evaluated on the BIRD-Mini Text-to-SQL benchmark (500 queries), Phebe achieves 61.6% execution accuracy - an 8 percentage point improvement over the GPT baseline - with a 0% SQL syntax error rate attributable to the self-correction loop. A user survey further confirms that 85% of participants (n = 27) rated their ability to explore complex datasets without writing SQL or code at 4 or 5 out of 5, validating the platform's goal of democratizing data exploration.