基于大模型语义抽取的多业态智能经营系统设计与实现

Design and Implementation of a Multi-Business Intelligent Operation System Based on Large Language Model Semantic Extraction

摘要: 针对中小企业经营数据管理过程中长期依赖Excel表格、数据结构不统一、跨业务数据整合困难以及智能化分析能力不足等问题，设计并实现了一种基于大语言模型语义对齐的多业态智能经营系统。系统采用数据录入与业务展示分离的架构，以PostgreSQL数据库作为统一数据源，利用开源Univer在线表格引擎提供低学习成本的数据录入能力，并通过大语言模型对表格字段语义进行理解与映射，实现不同业务模板下经营数据的自动化抽取与规范化入库。针对传统位置映射方式对表格结构变化适应性差的问题，提出基于表头语义匹配的数据对齐方法，并结合规则解析、模型辅助解析及指标约束校验机制，提高数据入库的稳定性与准确性。在此基础上，系统实现了协同数据录入、经营数据可视化、智能分析、AI 海报生成以及基于角色的权限管理等功能。通过采集足浴、酒店、月子中心和调理馆等多种业务场景的真实经营数据进行验证，其中足浴门店单月经营数据已完成实际部署应用。实验与应用结果表明，该系统能够有效降低中小企业数字化管理门槛，大语言模型可作为经营数据语义转换与结构化入库的重要支撑，为多业态企业智能化经营系统建设提供了一种可复用的实现方案。

Abstract: To address the challenges faced by small and medium-sized enterprises (SMEs), including heavy reliance on Excel-based management, inconsistent data structures, difficulties in cross-business integration, and insufficient intelligent analysis capabilities, this paper designs and implements a multi-business intelligent operation system based on large language model semantic alignment. The system adopts a separated architecture between data entry and business visualization, where PostgreSQL serves as the unified data source and the open-source Univer spreadsheet engine provides a low-cost data entry interface. A large language model is introduced to understand and align the semantics of spreadsheet fields, enabling automated extraction and standardized storage of operational data across different business templates. To overcome the limitations of traditional position-based mapping methods under changing spreadsheet layouts, a semantic matching approach based on table headers is proposed, combined with rule-based parsing, model-assisted extraction, and metric constraint validation mechanisms to improve the reliability and accuracy of data warehousing. Furthermore, the system implements collaborative data entry, operational dashboards, intelligent analysis, AI poster generation, and role-based access control. Real operational data from multiple business scenarios, including foot massage stores, hotels, maternity centers, and conditioning clinics, are collected for validation, and monthly operational data from a foot massage store has been successfully deployed in practice. Experimental and application results demonstrate that the proposed system effectively reduces the digital transformation barrier for SMEs. Large language models can serve as an effective semantic conversion layer for structured operational data management, providing a reusable solution for intelligent operation systems in multi-business enterprises.