Framework de validação semântica de ordens de manutenção com LLM e RAG

The semantic validation of maintenance orders and notes is an important element to ensure the integrity of the information that supports technical and strategic decisions in the mining industry. The lack of conceptual and semantic accuracy in these records can compromise not only the reliability of assets, but also performance indicators, maintenance plans and continuous improvement processes. In this scenario, the development of automated and intelligent tools that allow auditing, interpreting and justifying this data autonomously becomes a necessary and strategic response to the growing volume, complexity and criticality of operational infor- mation. This work proposes and develops a framework for the automated semantic validation of maintenance orders and notes, using Large Language Models (LLMs) or Large Language Models, executed locally, supported by a Retrieval-Augmented Generation (RAG) architecture. The system, implemented in Python, uses the Gemma-3:4B model via Ollama, ensuring data privacy, and uses the LangChain library to orchestrate the interaction. Custom knowledge bases extracted from business rules data are queried to provide domain-specific context to the LLM. Prompt engineering techniques such as Chain of Thought (CoT) that force the LLM to generate reasoning to promote explainability, as well as conversational memory management to optimize efficiency and reduce token processing, were used. This study contributes with a working proto- type, a systematic methodology for developing AI agents that are experts in semantic compliance analysis, aiming to improve data quality and support more accurate decisions in maintenance management.