libambu/data-agent

Data Agent is an end-to-end data analytics platform where you type a question in plain language and the system produces working SQL or Python code plus a Markdown analysis report. The project is built in Java and Vue 3, and its core is a network of AI agents that collaborate to answer your question. The pipeline is structured as a 13-node, 4-stage graph. When you submit a question, the system first retrieves relevant database schema information and business terminology from a vector store. It then evaluates whether the question is feasible, plans an execution strategy, and routes the work to specialized sub-agents: one that writes SQL queries, one that runs Python analysis or generates charts, and one that composes the final report. An LLM supervisor coordinates the sub-agents and can loop them if a result looks wrong. The system also includes a human-in-the-loop checkpoint where you can approve or reject the plan before execution proceeds. Data Agent uses a three-layer knowledge system backed by PostgreSQL with the pgvector extension. It stores database schema, business glossary terms, and historical question-answer pairs as vector embeddings, which it retrieves to guide each new query. The BIRD-SQL benchmark dataset (11 databases, 1,534 questions) is included for testing. Setting up the project requires Docker for PostgreSQL plus pgvector, Java 17, Maven, Node.js with pnpm, and a DashScope API key (an Alibaba Cloud service that provides access to the DeepSeek model used here). A Docker Compose file handles the database setup automatically. The frontend runs on Vue 3 with Element Plus and shows a live animated view of the agent graph as each node executes. The project is aimed at teams that want to give non-technical stakeholders a way to query databases by asking questions in plain text.