Graph-Augmented Retrieval Framework For Explainable Domain-Specific Large Language Model Reasoning

Abstract

Large language models have demonstrated strong capabilities in natural language understanding and reasoning; however, their use in domain-specific applications is limited by hallucination, weak factual grounding, and insufficient explainability. This study proposes a graph-augmented retrieval framework that combines semantic vector retrieval with knowledge-graph traversal to support accurate and transparent domain-specific reasoning. The framework retrieves relevant document segments, identifies connected entities and relationships, ranks the combined evidence, and supplies the selected knowledge to the language model for answer and explanation generation. The proposed method is evaluated against a standalone LLM, conventional retrieval-augmented generation, graph-only retrieval, and an existing GraphRAG approach. Performance is assessed using retrieval precision, answer accuracy, explainability score, hallucination rate, and response latency. The results demonstrate that the proposed framework improves evidence relevance, reasoning accuracy, and explanation quality while reducing unsupported responses, with a moderate increase in computational latency. The framework provides a practical foundation for trustworthy LLM deployment in knowledge-intensive domains.