This work presents an AI-enabled framework for evidence-based decision support in nuclear reactor licensing, addressing challenges in drafting safety analysis reports under regulatory guidance such as NUREG-1537. Licensing requires strict traceability, consistency across heterogeneous document sets, and auditable use of evidence—requirements that are difficult to satisfy with conventional large language model (LLM) approaches, which may produce unsupported or non-verifiable outputs, introducing risk in safety-critical documentation.
To mitigate these risks, we propose a retrieval-first, graph-based architecture that integrates semantic search with structured system knowledge. The framework combines a vector database for fine-grained evidence retrieval with a knowledge graph capturing document structure and model-based systems engineering (MBSE) relationships. An orchestration layer fuses these sources into a structured “ContextPack,” constraining LLM generation to traceable, verifiable evidence.
Controlled generation is enforced through multi-layer prompting, enabling citation fidelity, explicit gap identification, and completeness assessment. Hybrid retrieval strategies combining semantic similarity and keyword-based methods improve robustness and reduce the likelihood of missing critical information.
A use case involving research reactor licensing demonstrates how the framework enhances traceability, supports verification of completeness, and reduces the risk of undocumented or unsupported claims.