Recuperação de informações de um sistema ERP por meio da interpretação de linguagem natural utilizando RAG e banco de dados em grafos

Abstract

The effectiveness of business decision-making is directly related to the availability of reliable and well-structured information. In this context, Enterprise Resource Planning (ERP) systems play a fundamental role by centralizing large volumes of strategic data, facilitating their access and analysis. However, the use of these systems still represents a significant challenge for non-technical users, due to the need for knowledge of formal query languages and database-related concepts. This scenario reveals a gap in how data is accessed, making it difficult to retrieve information and support decision-making. In this context, this work presents the development of a prototype based on the Retrieval Augmented Generation (RAG) architecture, capable of interpreting natural language queries through the integration of language models and external data sources. The methodology involved modeling a simulated ERP scenario in the Neo4j graph database, defining a service-oriented architecture, and implementing a RAG orchestrator responsible for interpreting natural language queries, generating structured queries, and producing graphical responses based on the retrieved data, with the support of the LangChain library for integration with language models. The results demonstrated that the system was able to generate context-appropriate responses, with automatic presentation in the most suitable format. Additionally, the use of semantic caching contributed to performance optimization, reducing execution time and computational costs. The prototype contributes by expanding access to information in ERP systems, reducing the dependence on technical knowledge for their use, and making data analysis more accurate, positively impacting organizational efficiency and decision-making agility