In corporate activities, many business operations rely on internal industrial documents, such as laws/regulations, standards, specifications, design standards, and manufacturing instructions, as their underlying evidence for decision-making. These documents require referencing across hierarchical structures, citation relationships, defined terms and relevant documents. As a result, conventional retrieval-augmented generation (RAG) may fail to retrieve necessary supporting evidence. This report presents a method that uses a hybrid knowledge graph, combining structural and conceptual relationships, to expand candidate references using initial search results obtained via traditional RAG as a starting point. Moreover, large language models are used to extract candidate node types and relationship types tailored to the target document set, thereby supporting the initial development of the ontology design. Verification conducted using test data designed for internal workflows confirmed improvements in both the rate of relevant document retrieval and answer accuracy across high-pressure gas-related laws/regulations, Foreign Exchange and Foreign Trade Act-related laws/regulations, and the EU Artificial Intelligence Act.
