Local LLM Based Knowledge Graph in Power Domain Using Graph-Schema Reasoning and Retrieval-Augmented Generation

In recent years, large language models (LLMs) have shown great potential in many domains. However, their application in professional domains is often limited by problems like erroneous outputs and hallucinatory responses. Therefore, we present a framework that combines knowledge graphs (KGs) with local LLMs. The framework utilizes the factual information in KGs to improve the initial output of the LLMs, thereby reducing the factual errors in inference. In this paper, a domain knowledge graph is automatically constructed using textual data from the power industry. The KG contains 149,732 entities and 139,280 relationships. The proposed method is tested on EleQA, a public Q&A dataset of electricity regulations. Compared with the LLM-only baseline, the knowledge-graph-enhanced model achieves an improvement of 32.42%. Moreover, the framework shows strong adaptability and performs well on various LLMs. Our framework improves the accuracy and utility of large language models in the power domain.