Desenvolvimento de um sistema de IA generativa baseado em RAG para apoio à decisão na manutenção de equipamentos industriais

Abstract

This work presents the development of a prototype based on the Retrieval Augmented Generation (RAG) architecture aimed at optimizing the consultation of tech nical documentation in industrial maintenance environments. The manufacturing sector faces growing challenges in managing vast volumes of manuals and operational procedu res, a scenario where the rapid and accurate retrieval of information is critical for safety and the mitigation of equipment downtime. The proposed solution utilizes a Python bac kend structured with FastAPI and orchestrated by the LangChain library. The system integrates local embedding models (all-MiniLM-L6-v2) and the Chroma vector database to ensure source traceability. Response generation is driven by the Gemini model (Vertex AI), configured with strict temperature parameters and restrictive prompt engineering to suppress hallucinations. The user interface was built in Angular, prioritizing usability on the factory floor. System validation was performed through qualitative manual inspec tion, using a real industrial robotics manual from the manufacturer FANUC. Test results attest that the system is capable of extracting direct technical parameters, diagnosing error codes (troubleshooting), structuring step-by-step procedures, and interpreting sa fety restrictions with high factual accuracy. Furthermore, the prototype demonstrated remarkable resilience against adversarial tests and false positives, refraining from genera ting information when the documentary context was insufficient. It is concluded that the designed RAG architecture is technically viable, traceable, and safe to act as a decision support tool in complex industrial domains.