Abstract: The treatment of diabetes is a long-term and highly personalized endeavor and imposes a significant burden on patients’ daily lives.Diabetes consultation through medical large language models(LLMs) can effectively alleviate the medical healthcare burden of patients.But LLMs are more likely to produce hallucinations,i.e.,outputs that are incorrect,meaningless,or mismatched with the input,when processing texts in specialized domains such as medicine.And the accuracy rate of existing hallucination relief techniques in the medical field is not satisfactory,which will greatly affect the accuracy rate of the LLMs.To address this problem,this paper proposes a hallucination self-inspection and proactive relief method that combines instruction fine-tuning and retrieval augmented generation to form additional knowledge about user questions before the generation process,and to determine whether a hallucination is generated by similarity comparison after the generation process.Experiments are conducted on several medical datasets,and an F1 value of 0.79,a BLEU-4 value of 2.38,and a Rouge-l value of 9.26 are achieved on the large-scale diabetic multi-round conversation dataset,which outperforms the existing hallucination relief techniques for LLMs in terms of accuracy and generation efficiency.

Key words: Large language model, Retrieval augmented generation, Hallucination relief, Diabetes, Question and answer system