RADIANT-LLM: an Agentic Retrieval Augmented Generation Framework for Reliable Decision Support in Safety-Critical Nuclear Engineering

Researchers from Texas A&M University and Argonne National Laboratory have introduced RADIANT-LLM (Retrieval-Augmented, Domain-Intelligent Agent for Nuclear Technologies using LLM), a multi-modal RAG framework tailored for safety-critical nuclear engineering. The system addresses a persistent problem: LLMs hallucinate when handling specialized nuclear domain knowledge, often producing inaccurate or untraceable responses. RADIANT-LLM pairs a local-first, model-agnostic architecture with a structured, metadata-rich knowledge base, enabling page- and figure-level retrieval from technical documents like Used Nuclear Fuel Storage Facility design guidance.

The framework's agentic layer coordinates domain-specific tools, enforces citation-backed responses with full provenance tracking, and supports human-in-the-loop validation to further reduce hallucination risks. In rigorous evaluations using expert-curated benchmarks, RADIANT-LLM achieved Context Precision and Visual Recall scores within an 85-98% band, while hallucination rates were substantially lower than those observed in general-purpose LLM deployments. When the same queries were posed to commercial platforms without the RAG layer, hallucinations and citation errors increased markedly. These results underscore that locally controlled, multi-modal RAG with provenance enforcement is necessary for the factual accuracy and auditability that nuclear engineering workflows demand.