Autonomous SQL Agent: A Natural Language Database Interaction System Based on Multi-Agent Architecture

Autonomous SQL Agent is an enterprise-level multi-agent system designed to break down barriers for non-technical personnel to interact with databases. It connects natural language, real databases, and unstructured policy documents based on the LangGraph architecture, enabling secure database querying and intelligent Q&A. Core features include multi-agent collaboration, integration of Retrieval-Augmented Generation (RAG) technology, and strict safety guardrail mechanisms, providing an intelligent and secure data interaction solution for enterprise digital transformation.

In a data-driven business environment, non-technical personnel need to master SQL syntax to interact with databases. Existing natural language-to-SQL solutions have issues such as insufficient accuracy and lack of security. The Autonomous SQL Agent project emerged to build an intelligent and secure enterprise-level multi-agent system, allowing business personnel to directly interact with complex databases using natural language. This project adopts the LangGraph state machine architecture, upgrading the LangChain React loop into an extensible multi-agent system to achieve end-to-end natural language-to-SQL conversion, and integrates RAG to process unstructured policy documents.

Multi-Agent Collaboration Mechanism

The core of the system is an agent routing network, including a supervisory router (central coordinator), SQL expert agent (generates secure SQL queries), RAG expert agent (handles policy queries using ChromaDB for semantic mapping), and chat agent (handles non-database conversations).

Safety Execution Guardrails

The SQL engine intercepts generated queries, detects destructive operations (DROP/DELETE/UPDATE/INSERT), and ensures data security.

Core Technology Stack

LangGraph (multi-agent workflow), ChromaDB (vector storage), sentence-transformers (text embedding), Google Gemini/Hugging Face (AI models), Streamlit (web interface), SQLite+Pandas (data storage and analysis).

Dataset Validation

Using the Ames Housing Dataset (2930 rows, 82 columns) for demonstration, the system can automatically perform schema introspection and generate accurate aggregate queries.

Typical Scenarios

• Database query: For example, "What is the average selling price of houses with excellent basement quality?" generates aggregate SQL;
• Policy query: For example, "Can accessory dwelling units be built in RM zones?" retrieves policies via RAG;
• Mixed query: For example, "What is the average lot area in low-density residential areas?" first uses RAG for semantic mapping then executes SQL;
• Safety test: Intercepts destructive operations like "Delete the properties table".

Autonomous SQL Agent demonstrates the combination of LLM reasoning capabilities and structured data querying. It ensures reliability and security through multi-agent architecture and safety guardrails. It provides a reference implementation for intelligent agent workflows in the data engineering field, proving the feasibility of natural language interfaces in enterprise database query scenarios. For enterprises, it lowers the threshold for data access, enhances business personnel's self-service analysis capabilities, and offers an implementable technical solution.

In the future, we can further expand support for more database types, integrate enterprise knowledge bases, and add more complex permission control mechanisms to enhance the system's applicability and security, better serving the diverse needs of enterprises.