Medical Q&A Model Integrating Knowledge Graph and Vector Database with Two-Stage Retrieval and Feedback Optimization

With the rapid advancement of AI, the application potential of large language models (LLMs) in the medical field has become increasingly prominent. However, general-purpose LLMs still suffer from insufficient professional knowledge and a tendency to generate hallucinations in medical domains. Focusing on the core task of constructing a disease-oriented medical question-answering assistant, this paper proposes a medical intelligent question-answering system named HQ-LLM (HealthQuery Large Language Model) based on a two-stage retrieval and feedback iteration mechanism. Multi-source data from medical encyclopedias and public health authorities are integrated to construct a Chinese Medical Knowledge Graph and Question-Answering Dataset (CnMedKG-QA). The LoRA (Low-Rank Adaptation) technique is adopted to implement lightweight fine-tuning of the ChatGLM3-6B model, which can be supported by a single 32 GB vGPU for the training process, balancing cost control and training efficiency. To address the limitations of traditional RAG (Retrieval-Augmented Generation) systems, the paraphrase-multilingual-MiniLM-L12-v2 model is utilized for text vectorization to build a vector database with high relevance and retrieval efficiency. An innovative two-stage retrieval mechanism combining coarse-grained filtering and fine-grained reasoning is designed, along with a “retrieval-generation-evaluation-re-retrieval” feedback iteration mechanism. A medical-specific confidence evaluation system covering dimensions such as professional terminology accuracy and logical coherence is established. When the comprehensive confidence is below the threshold, retrieval key words are dynamically optimized, the number of knowledge graph tracing layers is expanded, or external authoritative data is supplemented to continuously improve retrieval quality. Experimental results show that the HQ-LLM achieves 20.47%, 42.12%, 50.26% and 79.20% in BLEU-4, Rouge-L, F1 Score and Accuracy, respectively, significantly outperforming existing medical models such as Qwen-UMLS-7B-Instruct.