This study presents a novel approach to knowledge graph generation from Japanese historical narratives using large language models through prompt-based and tool-based methodologies. The research introduces a comprehensive evaluation framework incorporating both supervised and unsupervised assessment techniques. A link prediction measure evaluates the internal consistency of the generated knowledge graphs, while a normalized graph edit distance metric quantifies global quality against gold standards. The framework aims to enhance AI reasoning capabilities through knowledge graph-based retrieval augmented generation, providing a robust foundation for structured knowledge representation and efficient information retrieval.