Large language models (LLMs) demonstrate strong general language capabilities but struggle with historical question answering due to fragmented reasoning and imprecise knowledge retrieval. To address these limitations, this study integrates finetuning, prompt engineering, and knowledge graph augmentation to enhance reasoning accuracy and factual consistency in complex historical queries. We fine-tune LLMs on structured historical datasets using multiple-choice formats to improve causal reasoning and temporal understanding. Chain-of-thought prompting strategies, including zero-shot and few-shot paradigms, guide models to systematically decompose questions and synthesize logical answers. A domain-specific knowledge graph supplements LLMs with structured historical relationships, enabling real-time fact verification and reducing hallucinations. Experimental results show significant improvements in both answer accuracy and contextual depth, validating the effectiveness of combining these techniques. The framework demonstrates transferability to other specialized domains, offering a versatile approach to enhance reliability in expert-level question answering while maintaining the flexibility of pre-trained language models.