【Mini AI Agent】如何用 100 行代码构建一个最小智能体？

from langchain_community.llms import QianfanLLMEndpoint
import os

print("# 初始化千帆")
os.environ["QIANFAN_AK"]='千帆应用的 API Key'
os.environ["QIANFAN_SK"]='千帆应用的 Secret Key'

llm = QianfanLLMEndpoint(streaming=True,
                         model="ERNIE-3.5-8K",
                         temperature=0.1)

response = llm.invoke("上海春天一般哪个月开始？")
print(response)

from langchain import PromptTemplate, LLMChain
from langchain.chains import LLMRequestsChain
from langchain_core.tools import tool
from langchain_core.utils.function_calling import convert_to_openai_tool

print("# 定义函数")
@tool
def search_ip(question:str, ip:str) -> str:
    """
    首先获取输入的IP地址的位置信息，然后回答输入的问题
    @param question: 问题
    @param ip: IP地址
    """
    prompt_template = """以下是IP地址'{ip}'的位置信息：
    >>> {requests_result} <<<
    根据以上位置信息，回答以下这个问题：
    >>> {question} <<<"""
    prompt = PromptTemplate(
        input_variables=["question", "ip", "requests_result"],
        template=prompt_template
    )
    chain = LLMRequestsChain(llm_chain = LLMChain(llm=llm, prompt=prompt))
    inputs = {
        "question": question,
        "ip": ip,
        "url": "https://api.songzixian.com/api/ip?dataSource=generic_ip&ip=" + ip
    }
    return chain.invoke(inputs)

@tool
def search_phone(question:str, phone:str) -> str:
    """
    首先获取输入的手机号码的归属地信息，然后回答输入的问题
    @param question: 问题
    @param phone: 手机号码
    """
    prompt_template = """以下是手机号码'{phone}'的归属地信息：
    >>> {requests_result} <<<
    根据以上归属地信息，回答以下这个问题：
    >>> {question} <<<"""
    prompt = PromptTemplate(
        input_variables=["question", "phone", "requests_result"],
        template=prompt_template
    )
    chain = LLMRequestsChain(llm_chain = LLMChain(llm=llm, prompt=prompt))
    inputs = {
        "question": question,
        "phone": phone,
        "url": "https://api.songzixian.com/api/phone-location?dataSource=phone_number_location&phoneNumber=" + phone
    }
    return chain.invoke(inputs)

functions=[convert_to_openai_tool(search_ip)['function'],convert_to_openai_tool(search_phone)['function']]
print(functions)

from langchain.schema import HumanMessage
from langchain_community.chat_models import QianfanChatEndpoint
import json

# 此处不能设定streaming=True，否则无法激活函数回调
chat = QianfanChatEndpoint(model="ERNIE-3.5-8K",
                          temperature=0.1)

def get_response(message):
    print("# 绑定函数")
    result = chat.invoke([HumanMessage(content=message)],
                               functions=functions)
    print(result)

    function_call_info = result.additional_kwargs.get("function_call", None)
    print(function_call_info)
    
    if not function_call_info:
        print("# 直接返回")
        return result.content

    print("# 调用函数")
    function_name = function_call_info["name"]
    function_args = json.loads(function_call_info["arguments"])
    function_result = eval(function_name)(function_args)
    print(function_result)
    return function_result["output"]

import gradio as gr

def submit(message, chat_history):
    bot_message = get_response(message)
    # 保存历史对话记录，用于显示
    chat_history.append((message, bot_message))
    return "", chat_history

print("# 创建交互")
with gr.Blocks() as demo:
    chatbot = gr.Chatbot(height=240) # 对话框
    msg = gr.Textbox(label="Prompt") # 输入框
    submitBtn = gr.Button("Submit") # 提交按钮
    clearBtn = gr.ClearButton([msg, chatbot]) # 清除按钮
    # 提交
    msg.submit(submit, inputs=[msg, chatbot], outputs=[msg, chatbot]) 
    submitBtn.click(submit, inputs=[msg, chatbot], outputs=[msg, chatbot])
    
gr.close_all()
demo.launch()