如何用大模型RAG做医疗问答系统

代码参考

• https://github.com/honeyandme/RAGQnASystem
• https://github.com/LongxingTan/open-retrievals
• https://github.com/WangRongsheng/MedQA-ChatGLM
• https://github.com/baiyang2464/chatbot-base-on-Knowledge-Graph

TLDR

if '疾病症状' in entities and  '疾病' not in entities:
        sql_q = "match (a:疾病)-[r:疾病的症状]->(b:疾病症状 {名称:'%s'}) return a.名称" % (entities['疾病症状'])
        res = list(client.run(sql_q).data()[0].values())
        # print('res=',res)
        if len(res)>0:
            entities['疾病'] = random.choice(res)
            all_en = "、".join(res)
            prompt+=f"<提示>用户有{entities['疾病症状']}的情况，知识库推测其可能是得了{all_en}。请注意这只是一个推测，你需要明确告知用户这一点。"

根据实体确定图数据库查询语句，从中查询得到结果。疾病症状和知识库查询结果一起组成prompt，输入大模型中

系统设计

实体识别

• token classification

意图识别

• sequence classification

知识图谱

• graph

对话

搭建知识图谱

实体识别

LLM做的特点

优势

• they can handle a broad spectrum of entity types;
• they are highly adaptable to various domains and languages;
• their performance often surpasses that of traditional rule-based (e.g. regular expressions) or feature-based NER system;
• they can capture contextual information and context dependencies more effectively (e.g. sentiment analysis or intent detection);
• LLMs are capable of transfer learning, meaning they can be pre-trained on a general language corpus and fine-tuned for specific NER tasks, thus requiring fewer annotated data points for training.

con

• LLMs may raise concerns about model bias, model interpretability and ethical considerations, which require careful attention;
• LLM responses may contain “hallucinations” that can lead to the spread of misinformation;
• Fine-tuning requires designing appropriate training data, carefully selecting hyper-parameters, and often involves substantial computational resources.

prompt

intent
intent_name_field = ResponseSchema(name=“intent”, description=f"Based on the latest user message, extract the user message intent. Here are some possible labels: ‘greetings’, ‘booking’, ‘complaint’ or ‘other’")

user need
user_need_field = ResponseSchema(name=“user_need”, description=“Rephrase the latest user request and make it a meaningful question without missing any details. Use ‘’ if it is not available”)

user sentiment
sentiment_field = ResponseSchema(name=“sentiment”, description=“Based on the latest user message, extract the user sentiment. Here are some possible labels: ‘positive’, ‘neutral’, ‘negative’, ‘mixed’ or ‘other’”)

number of pizzas to be ordered
n_pizzas_field = ResponseSchema(name=“n_pizzas”, description=“Based on the user need, extract the number of pizzas to be made. Use ‘’ if it is not available”)