PV-RAG：基于LLM的药物警戒检索增强生成模型

PV-RAG: Retrieval-Augmented Generation Model for Pharmacovigilance Based on Large Language Models

摘要: 目的/意义：随着大语言模型(LLM)在自然语言处理任务中的广泛应用，其在处理专业领域问答任务中表现出局限性。药品因其“治病且可能致病”的双重特征，须在安全警戒范围内进行使用，而传统大模型受训练数据影响难以有效表达用药安全复杂且动态的信息，易导致数据缺失引发的回答幻觉问题。方法/过程：本文提出了一种面向药物警戒的检索增强生成模型(PV-RAG)，其综合应用大语言模型LangChain框架、稀疏与稠密检索混合策略以及改进的查询改写机制，显著提升检索生成结果的准确性。结果/结论：对比实验结果表明，PV-RAG模型能够精准提取药品说明书中的关键信息，在药物警戒问答任务中表现出显著优势，为药物警戒智能化问答提供了新的技术支持和应用前景。

Abstract: Purpose/Significance: With the widespread application of large language models (LLMs) in natural language processing tasks, their limitations in handling domain-specific question-answering tasks have become evident. Pharmaceuticals, characterized by their dual nature of “treating diseases while potentially causing harm”, must be used within strict safety monitoring frameworks. Traditional large models, constrained by training data, struggle to accurately represent the complex and dynamic nature of drug safety information, often leading to hallucinated responses due to data gaps. Methods/Process: This study proposes a Pharmacovigilance-oriented Retrieval-Augmented Generation model (PV-RAG), which integrates the LangChain framework for LLMs, a hybrid sparse-dense retrieval strategy, and an improved query rewriting mechanism to significantly enhance the accuracy of retrieval-generated responses. Results/Conclusions: Comparative experiments demonstrate that the PV-RAG model can precisely extract critical information from drug manuals and exhibits superior performance in pharmacovigilance question-answering tasks. This approach provides new technical support and application prospects for intelligent pharmacovigilance systems.