Research and Practice of NL2SQL Technology Based on LLM for Big Data of Enterprise Finance

Currently, the research field of Natural Language Generated SQL (NL2SQL) mainly focuses on generic datasets, aiming to build models that can parse natural language queries and automatically generate SQL statements. However, this generic exploration often ignores the complexity and idiosyncrasies of intra-enterprise data, such as industry-specific terminology, data structure differences, and security compliance requirements, resulting in the fact that the existing NL2SQL technology can only cover the more basic query requirements in practical applications, and is difficult to be deeply integrated into the business scenarios of enterprises. This paper aims to fill this research gap by focusing on the customized application of NL2SQL technology to specific internal enterprise environments, and adopting LLM+RAG+Reminiscence Engineering+Intelligent Body Feedback to design and implement a set of NL2SQL system that can be applied to internal enterprises. After practical project verification, the system’s ability to generate SQL in an enterprise-oriented financial data environment can be improved from 54% to about 70%, and the accuracy of multi-round dialog can be further improved. This system enables the seamless connection between natural language and enterprise database, which provides strong support for enterprise digital transformation.