RADIANT-LLM: A Framework for Enhanced Decision Support in Nuclear Engineering

In the field of nuclear engineering, effective decision-making hinges on dependable access to domain-specific knowledge. However, existing workflows often struggle with disorganized documentation and inaccuracies when utilizing pre-trained large language models (LLMs). To tackle these issues, we introduce RADIANT-LLM, a Retrieval-Augmented Generation (RAG) framework tailored for applications in nuclear safety, security, and safeguards. This innovative framework employs a local-first, model-agnostic design, integrating a multi-modal document ingestion system with a comprehensive, metadata-rich knowledge base that allows for detailed retrieval from technical documents. An agentic layer is incorporated to manage domain-specific tools, ensuring responses are citation-backed and traceable, while also facilitating human oversight to minimize inaccuracies. To assess the effectiveness of RADIANT-LLM, we developed a set of specialized metrics such as Context Precision (CoP) and Hallucination Rate (HR), tested against expert-curated benchmarks. Results demonstrate that CoP and Visual Recall (ViR) consistently remain high, while hallucination rates are significantly lower compared to standard LLM applications. These findings underscore the necessity of a structured, multi-modal RAG framework to meet the stringent accuracy and transparency requirements of nuclear engineering.