Diseño y desarrollo de un agente experto basado en LLM para automatizar consultas y facilitar la toma de decisiones en la gestión de Ventas, Compras e Inventarios de la Empresa Lojagas C.E.M.

Abstract

Abstract: Currently, the design and development of an expert agent based on large-scale language models plays a fundamental role in automating queries and decision-making at the enterprise level. This study addresses the management of sales, purchases, and inventory at Lojagas company, a solution that integrates Fine-Tuning and RAG techniques on Odoo ERP, deployed on local servers to ensure the confidentiality of sensitive data. This agent was implemented using modular architecture based on ISO 42010 and Kruchten's 4+1 views model. The Mistral model optimized through QLoRA and ChromaDB as a vector database were used. Furthermore, mixed validation demonstrated 94% accuracy in contextual fidelity, 91% relevance, and a SUS score of 90.5 in usability, with average response times of 12.5 seconds, showing that the results confirm the feasibility of implementing specialized conversational agents on low-performance equipment, significantly improving operational efficiency and user experience in business contexts, without compromising information security.