A retrieval-augmented, domain-intelligent agentic framework for reliable decision support in safety-critical nuclear engineering

Reliable decision support in nuclear engineering requires traceable retrieval from technical documentation, yet conventional deployment of pretrained large language models (LLMs) in specialized workflows remains vulnerable to fragmented source access, weak traceability, and hallucination. This paper presents RADIANT-LLM (Retrieval-Augmented, Domain-Intelligent Assistant for Nuclear Technologies using LLMs), a training-free, local-first multimodal visual retrieval-augmented generation (Visual-RAG) framework for nuclear safety, security, and safeguards applications. The framework integrates multimodal PDF ingestion, structured local knowledge-base construction, metadata-aware retrieval, and agentic tool orchestration to answer questions over documents containing text, equations, tables, figures, and schematics. By retaining documents, parsed content, embeddings, and knowledge-base assets on user-controlled infrastructure, and transmitting only retrieved evidence when external models are used, the framework is designed to reduce broad document exposure while preserving page- and figure-level traceability. The framework was evaluated using expert-curated benchmarks derived from authoritative technical PDFs, including guidance for Used Nuclear Fuel Storage Facility (UNFSF) design, and five domain-aware metrics: Context Precision (CoP), Citation Precision (CiP), Citation Hit (CiH), Hallucination Rate (HR), and Visual Recall (ViR). A two-stage evaluation first identified a strong multimodal RAG configuration and then assessed performance as the knowledge base expanded from 1 to 250 sources. Across corpus expansion, retrieval quality and visual grounding remained strong (CoP: 0.813–0.956; CiP: 0.893–0.964; CiH: 0.900–0.989; ViR: 0.802–0.983), while hallucination remained low (HR: 0.024–0.094), with no statistically significant monotonic degradation. Compared with general-purpose LLM use without a dedicated RAG layer, RADIANT-LLM produced more reliable, citation-aligned, page- and figure-grounded responses.