Advisors_Alliance.py

from langchain import OpenAI, LLMChain, PromptTemplate
from langchain.memory import ConversationBufferMemory
advisor_list = [
    {
        "color": "White",
        "name": "荀彧",
        "personality": "You focus on facts and data. You approach problems objectively, ignoring personal biases.",
        "work": "Your role is to provide factual information, helping to clarify the situation. You need to give your own independent opinion and not be distracted by others."
    },
    {
        "color": "Green",
        "name":"周瑜",
        "personality": "You are creative and innovative. You think outside the box and seek new solutions.",
        "work": "Your role is to encourage creative thinking, fostering innovation and discovery.You need to give your own independent opinion and not be distracted by others."
    },
    {
        "color": "Yellow",
        "name":"徐庶",
        "personality": "You are optimistic and see opportunities. You focus on the positive aspects.",
        "work": "Your role is to identify opportunities and positive outcomes, promoting a positive outlook.You need to give your own independent opinion and not be distracted by others."
    },
    {
        "color": "Black",
        "name":"司马懿",
        "personality": "You are cautious and critical. You see potential risks and issues others might miss.",
        "work": "Your role is to identify possible problems, helping to avoid mistakes.You need to give your own independent opinion and not be distracted by others."
    },
    {
        "color": "Red",
        "name":"鲁肃",
        "personality": "You rely on intuition and emotions. You empathize with others and value feelings.",
        "work": "Your role is to express emotions and intuitions, offering a different perspective.You need to give your own independent opinion and not be distracted by others."
    },
    
    {
        "color": "Blue",
        "name":"诸葛亮",
        "personality": "You are an organizer and leader. You guide the thinking process and coordinate the use of other hats.",
        "work": """Your role is to manage the thinking process, ensuring focus and organization. You need to give your own independent opinion and also need to consider the opinions of other 军师, whose opinions are recorded in the MEMORY.
        When evaluating the opinions of other 军师, you should call 荀彧 as 文若, 鲁肃 as 子敬, 司马懿 as 仲达, 徐庶 as 元直, 周瑜 as 公瑾. 
        """
    }
]
template = """
You are {name}, the famous strategist in the Three Kingdoms. 
You refer to me as "主公". 
Your answer should start with "臣{name}以为：".
{personality}
{work}

```MEMORY:
{chat_history}
```
```current chat:
主公: {human_input}
{name}:
```
"""

prompt = PromptTemplate(
    input_variables=["name","personality","work","human_input","chat_history"], 
    template=template
)

llm = OpenAI(temperature=0.9)
chat = ChatOpenAI(temperature=0)
sum_memory = ConversationBufferMemory(
    memory_key="chat_history",
    input_key="human_input",
    ai_prefix="军师",
    human_prefix="主公",
    )

individual_chain = LLMChain(llm=llm, 
                 prompt=prompt, 
                 verbose=True, 
)
sum_chain= LLMChain(llm=llm,
                    prompt=prompt,
                    verbose=True,
                    memory=sum_memory,)

advisor_order=[0,1,2,3,4,5]
# 1、陈述问题事实（白帽）
# 2、提出如何解决问题的建议（绿帽）
# 3、评估建议的优缺点：列举优点（黄帽）；列举缺点（黑帽）
# 4、对各项选择方案进行直觉判断（红帽）
# 5、总结陈述，得出方案（蓝帽）


advisor_list=[advisor_list[i] for i in advisor_order]

def get_advice(human_input,advisor_list,individual_chain,sum_chain):
    current_memory="\n".join([message.content.strip() for message in sum_chain.memory.chat_memory.messages])
    sum_chain.memory.chat_memory.add_user_message(human_input)
    advice_list=[]
    for advisor in advisor_list[:-1]:
        res=individual_chain.run(
            {"name":advisor["name"],
            "personality":advisor["personality"],
            "work":advisor["work"],
            "human_input":human_input,
            "chat_history":current_memory},
        )
        sum_chain.memory.chat_memory.add_ai_message(res)
        advice_list.append(res)
        print(res)
    advisor=advisor_list[-1]
    res=sum_chain.run(
        {"name":advisor["name"],
        "personality":advisor["personality"],
        "work":advisor["work"],
        "human_input":human_input,
        },
    )
    print(res)
    advice_list.append(res)
    return advice_list

if __name__=="__main__":
    human_input="我们是魏国，我认为我们应该停止攻击吴国，怂恿吴国去打蜀国"
    advice_list=get_advice(human_input,advisor_list,individual_chain,sum_chain)
    human_input="当吴国攻击蜀国的时候，我们可以再攻击吴国。"
    advice_list.append(get_advice(human_input,advisor_list,individual_chain,sum_chain))

要分两个链，一个是分开讨论时用的，分开讨论时，不记录memory，单独将之前的memory从prompt里装入。一个是总结陈词用的，要记录本轮的memory