Speaker
Description
In the nuclear energy sector, adherence to safety standards is paramount for operational efficiency and radiological protection. However, the exponential growth of technical documents issued by the International Atomic Energy Agency (IAEA) and the Brazilian National Commission for Nuclear Energy (CNEN) has made efficient information retrieval a critical challenge. Traditional keyword-based search methods often yield generic results, while the direct application of Large Language Models (LLMs) is limited by the risk of "hallucinations"—the generation of plausible but factually incorrect information. To address these limitations, this work proposes a system based on Retrieval-Augmented Generation (RAG). This methodology combines the generative capabilities of LLMs with a structured retrieval mechanism that accesses a specific knowledge base of nuclear regulations. The system utilizes semantic embedding strategies and vector databases to ground the model's responses in precise regulatory context. Furthermore, the project explores Low-Rank Adaptation (LoRA) to fine-tune models for the specific terminology of the nuclear domain without prohibitive computational costs. The primary objective is to ensure traceability and auditability; every response generated cites the specific document and section of the IAEA or CNEN standard. By providing a reliable assistant for interpreting complex dose limits, shielding requirements, and safety guides, this tool directly supports decision-making in computational dosimetry. Preliminary evaluations indicate that the RAG-based approach significantly improves the accuracy of regulatory interpretation compared to standard LLMs, enhancing the operational efficiency of radiation protection officers.
Keywords: Artificial Intelligence; Radiation Protection; Regulatory Compliance; Retrieval-Augmented Generation; IAEA Standards.