Title:SchemaAgent: A Multi-Agents Framework for Generating Relational Database Schema

Abstract:The relational database design would output a schema based on user's requirements, which defines table structures and their interrelated relations. Translating requirements into accurate schema involves several non-trivial subtasks demanding both database expertise and domain-specific knowledge. This poses unique challenges for automated design of relational databases. Existing efforts are mostly based on customized rules or conventional deep learning models, often producing suboptimal schema. Recently, large language models (LLMs) have significantly advanced intelligent application development across various domains. In this paper, we propose SchemaAgent, a unified LLM-based multi-agent framework for the automated generation of high-quality database schema. SchemaAgent is the first to apply LLMs for schema generation, which emulates the workflow of manual schema design by assigning specialized roles to agents and enabling effective collaboration to refine their respective subtasks. Schema generation is a streamlined workflow, where directly applying the multi-agent framework may cause compounding impact of errors. To address this, we incorporate dedicated roles for reflection and inspection, alongside an innovative error detection and correction mechanism to identify and rectify issues across various phases. For evaluation, we present a benchmark named \textit{RSchema}, which contains more than 500 pairs of requirement description and schema. Experimental results on this benchmark demonstrate the superiority of our approach over mainstream LLMs for relational database schema generation.