In recent years, large language models (LLMs) have shown great performance in natural language processing tasks, and they are also gradually being used in smart recommendation systems. However, since LLMs mainly rely on their parametric knowledge, they often face issues like delayed knowledge updates and hallucination problems when handling personalized recommendation tasks in specific fields. In addition, directly updating the internal parameters of LLMs to meet recommendation needs in specific fields usually comes with high computational costs. While traditional retrieval-augmented generation (RAG) methods can improve recommendations by retrieving external knowledge, they still have these limitations: Firstly, the model performs external searches for every recommendation, even when the LLMs’ own knowledge is sufficient. This redundant process significantly reduces system efficiency. Secondly, the quality of retrieved content directly affects recommendations. If the retrieved content is irrelevant or contains noise, the recommendation accuracy may drop. To solve these problems, this research proposes a education retrieval-augmented generation framework (EduRAG) specifically for academic expert recommendation tasks. EduRAG combines and improves GraphRAG and VectorRAG technologies to increase recommendation accuracy and reliability. The experiment results show EduRAG performs better than traditional RAG methods in both matching precision and recommendation quality for academic experts, offering an efficient and smart solution for academic resource matching.