欢迎光临
我们一直在努力

Day13:迈向智能交互:从 ReAct Agent 开发到优化的全历程

Day13:迈向智能交互:从 ReAct Agent 开发到优化的全历程

一、引言

在当今人工智能应用的浪潮中,开发智能 Agent 成为解锁大语言模型潜力的关键。本周我们将深入学习如何开发一个完整的 ReAct Agent,并在后续对其进行复盘、调试与优化。以通义千问大模型为依托,逐步构建一个功能强大、交互友好的智能 Agent。

二、第一个完整 ReAct Agent 开发

核心任务:整合知识开发完整 ReAct Agent

  • 工具选择与整合:我们需要整合多种工具,使 Agent 能够根据用户问题自主选择合适工具。例如,我们选择计算器、搜索和文件读取作为初始的三个工具。对于计算器工具,用于处理数值计算问题;搜索工具,可针对用户信息查询需求;文件读取工具,则能满足读取本地文件内容的要求。

  • 工具函数定义:

    • 计算器工具:
  • class CalculatorParams(BaseModel):
    num1: float
    num2: float
    operation: str

    @validator('operation')
    def valid_operation(cls, v):
    valid_ops = ['+', '-', '*', '/']
    if v not in valid_ops:
    raise ValueError('不支持的运算操作')
    return v

    def calculate(params: CalculatorParams):
    if params.operation == '+':
    return params.num1 + params.num2
    elif params.operation == '-':
    return params.num1 – params.num2
    elif params.operation == '*':
    return params.num1 * params.num2
    elif params.operation == '/':
    if params.num2 == 0:
    print("除数不能为零")
    return None
    return params.num1 / params.num2

    搜索工具:假设使用一个模拟的搜索 API,实际需替换为真实 API。

    class SearchParams(BaseModel):
    query: str
    api_key: str

    @validator('query')
    def query_not_empty(cls, v):
    if not v.strip():
    raise ValueError('查询内容不能为空')
    return v

    @validator('api_key')
    def api_key_not_empty(cls, v):
    if not v.strip():
    raise ValueError('api_key不能为空')
    return v

    def search(params: SearchParams):
    base_url = "https://api.example.com/search"
    params_dict = {
    "key": params.api_key,
    "q": params.query
    }
    try:
    response = requests.get(base_url, params=params_dict)
    response.raise_for_status()
    return response.json()
    except requests.RequestException as e:
    print(f"搜索失败: {e}")
    return None

    文件读取工具:

    class ReadFileParams(BaseModel):
    file_path: str

    @validator('file_path')
    def file_path_exists(cls, v):
    if not os.path.exists(v):
    raise ValueError('文件路径不存在')
    return v

    def read_file(params: ReadFileParams):
    try:
    with open(params.file_path, 'r', encoding='utf – 8') as f:
    return f.read()
    except Exception as e:
    print(f"读取文件失败: {e}")
    return None

    工具选择逻辑:借助通义千问大模型的推理能力,判断用户问题所需工具。通过向通义千问发送用户问题和工具描述,模型决定调用哪个工具。

    def choose_tool(user_input, api_key):
    tools = [
    {
    "name": "calculate",
    "parameters": {
    "type": "object",
    "properties": {
    "num1": {"type": "number"},
    "num2": {"type": "number"},
    "operation": {"type": "string"}
    },
    "required": ["num1", "num2", "operation"]
    },
    "description": "用于执行简单的数学运算"
    },
    {
    "name": "search",
    "parameters": {
    "type": "object",
    "properties": {
    "query": {"type": "string"},
    "api_key": {"type": "string"}
    },
    "required": ["query", "api_key"]
    },
    "description": "用于搜索信息"
    },
    {
    "name": "read_file",
    "parameters": {
    "type": "object",
    "properties": {
    "file_path": {"type": "string"}
    },
    "required": ["file_path"]
    },
    "description": "用于读取本地文件内容"
    }
    ]
    prompt = {
    "user_input": user_input,
    "functions": tools
    }
    response = call_qwen_api(api_key, json.dumps(prompt))
    if "function_call" in response:
    function_call = response["function_call"]
    return function_call["name"], function_call["parameters"]
    return None, None

    结果解析与回答生成:不同工具返回结果格式不同,需要针对性解析。例如,计算器返回简单数值,搜索可能返回 JSON 格式数据,文件读取返回文本内容。解析后,构建提示让通义千问生成自然语言回答。

    def parse_calculator_result(result):
    if result is not None:
    return f"计算结果为{result}"
    return ""

    def parse_search_result(result):
    if result:
    # 假设搜索结果结构,实际需根据真实API调整
    return f"搜索结果: {result.get('results', [])}"
    return ""

    def parse_read_file_result(result):
    if result:
    return f"文件内容: {result}"
    return ""

    def generate_answer(parsed_result, api_key):
    prompt = f"根据以下信息生成回答: {parsed_result}"
    response = call_qwen_api(api_key, prompt)
    if response:
    return response["choices"][0]["message"]["content"]
    return "无法生成回答"

    完整 ReAct Agent 代码:

    import requests
    import json
    import os
    from pydantic import BaseModel, validator

    # 计算器工具
    class CalculatorParams(BaseModel):
    num1: float
    num2: float
    operation: str

    @validator('operation')
    def valid_operation(cls, v):
    valid_ops = ['+', '-', '*', '/']
    if v not in valid_ops:
    raise ValueError('不支持的运算操作')
    return v

    def calculate(params: CalculatorParams):
    if params.operation == '+':
    return params.num1 + params.num2
    elif params.operation == '-':
    return params.num1 – params.num2
    elif params.operation == '*':
    return params.num1 * params.num2
    elif params.operation == '/':
    if params.num2 == 0:
    print("除数不能为零")
    return None
    return params.num1 / params.num2

    # 搜索工具
    class SearchParams(BaseModel):
    query: str
    api_key: str

    @validator('query')
    def query_not_empty(cls, v):
    if not v.strip():
    raise ValueError('查询内容不能为空')
    return v

    @validator('api_key')
    def api_key_not_empty(cls, v):
    if not v.strip():
    raise ValueError('api_key不能为空')
    return v

    def search(params: SearchParams):
    base_url = "https://api.example.com/search"
    params_dict = {
    "key": params.api_key,
    "q": params.query
    }
    try:
    response = requests.get(base_url, params=params_dict)
    response.raise_for_status()
    return response.json()
    except requests.RequestException as e:
    print(f"搜索失败: {e}")
    return None

    # 文件读取工具
    class ReadFileParams(BaseModel):
    file_path: str

    @validator('file_path')
    def file_path_exists(cls, v):
    if not os.path.exists(v):
    raise ValueError('文件路径不存在')
    return v

    def read_file(params: ReadFileParams):
    try:
    with open(params.file_path, 'r', encoding='utf – 8') as f:
    return f.read()
    except Exception as e:
    print(f"读取文件失败: {e}")
    return None

    # 模拟通义千问API调用(实际需替换为真实API调用)
    def call_qwen_api(api_key, prompt):
    url = "https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions"
    headers = {
    "Content-Type": "application/json",
    "Authorization": f"Bearer {api_key}"
    }
    data = {
    "model": "qwen-plus",
    "messages": [
    {"role": "user", "content": prompt}
    ],
    "stream": False
    }
    try:
    response = requests.post(url, headers=headers, json=data)
    response.raise_for_status()
    return response.json()
    except requests.RequestException as e:
    print(f"调用通义千问API失败: {e}")
    return None

    def choose_tool(user_input, api_key):
    tools = [
    {
    "name": "calculate",
    "parameters": {
    "type": "object",
    "properties": {
    "num1": {"type": "number"},
    "num2": {"type": "number"},
    "operation": {"type": "string"}
    },
    "required": ["num1", "num2", "operation"]
    },
    "description": "用于执行简单的数学运算"
    },
    {
    "name": "search",
    "parameters": {
    "type": "object",
    "properties": {
    "query": {"type": "string"},
    "api_key": {"type": "string"}
    },
    "required": ["query", "api_key"]
    },
    "description": "用于搜索信息"
    },
    {
    "name": "read_file",
    "parameters": {
    "type": "object",
    "properties": {
    "file_path": {"type": "string"}
    },
    "required": ["file_path"]
    },
    "description": "用于读取本地文件内容"
    }
    ]
    prompt = {
    "user_input": user_input,
    "functions": tools
    }
    response = call_qwen_api(api_key, json.dumps(prompt))
    if "function_call" in response:
    function_call = response["function_call"]
    return function_call["name"], function_call["parameters"]
    return None, None

    def parse_calculator_result(result):
    if result is not None:
    return f"计算结果为{result}"
    return ""

    def parse_search_result(result):
    if result:
    return f"搜索结果: {result.get('results', [])}"
    return ""

    def parse_read_file_result(result):
    if result:
    return f"文件内容: {result}"
    return ""

    def generate_answer(parsed_result, api_key):
    prompt = f"根据以下信息生成回答: {parsed_result}"
    response = call_qwen_api(api_key, prompt)
    if response:
    return response["choices"][0]["message"]["content"]
    return "无法生成回答"

    def react_agent(user_input, api_key):
    tool_name, tool_params = choose_tool(user_input, api_key)
    if tool_name == 'calculate':
    params = CalculatorParams(**tool_params)
    result = calculate(params)
    parsed_result = parse_calculator_result(result)
    elif tool_name =='search':
    params = SearchParams(**tool_params)
    result = search(params)
    parsed_result = parse_search_result(result)
    elif tool_name =='read_file':
    params = ReadFileParams(**tool_params)
    result = read_file(params)
    parsed_result = parse_read_file_result(result)
    else:
    return "无法确定要使用的工具"
    return generate_answer(parsed_result, api_key)

    补充任务:优化交互体验

  • 用户输入提示:在程序开始时,提示用户输入问题,使交互更友好。

  • 执行步骤打印:在 Agent 运行过程中,打印选择的工具、传入的参数以及中间结果,方便追溯问题。

  • if __name__ == "__main__":
    api_key = "your_api_key"
    user_input = input("请输入你的问题: ")
    print(f"用户输入: {user_input}")
    tool_name, tool_params = choose_tool(user_input, api_key)
    print(f"选择的工具: {tool_name}")
    print(f"工具参数: {tool_params}")
    if tool_name == 'calculate':
    params = CalculatorParams(**tool_params)
    result = calculate(params)
    print(f"计算结果: {result}")
    parsed_result = parse_calculator_result(result)
    elif tool_name =='search':
    params = SearchParams(**tool_params)
    result = search(params)
    print(f"搜索结果: {result}")
    parsed_result = parse_search_result(result)
    elif tool_name =='read_file':
    params = ReadFileParams(**tool_params)
    result = read_file(params)
    print(f"文件读取结果: {result}")
    parsed_result = parse_read_file_result(result)
    else:
    print("无法确定要使用的工具")
    exit()
    answer = generate_answer(parsed_result, api_key)
    print(f"生成的回答: {answer}")

    三、 Agent 调试优化

    核心任务:复盘本周学习内容

  • Function Call:回顾如何使用 Function Call 让大模型调用外部工具,包括工具函数定义、参数校验以及与大模型的交互流程。

  • ReAct 范式:总结 ReAct 范式的 “推理→动作→观察→再推理” 流程,以及它如何与 Function Call 结合,使 Agent 能够根据用户问题合理选择并调用工具。

  • Agent 开发:梳理开发完整 ReAct Agent 的过程,从工具选择、工具函数实现、结果解析到回答生成的各个环节,分析每个环节的作用和实现方法。

  • 补充任务:调试优化 Agent

  • 修复已知 bug:

    • 工具选择错误:检查工具选择逻辑,确保通义千问返回的工具选择与用户问题匹配。可能需要调整工具描述,使模型更准确理解工具适用场景。

    • 结果解析失败:针对不同工具返回结果,仔细检查解析函数。例如,若搜索 API 返回结果结构变化,相应调整 parse_search_result 函数。

  • 增加新工具(文件写入):

    • 工具函数定义:

      class WriteFileParams(BaseModel):
      file_path: str
      content: str

      @validator('file_path')
      def file_path_valid(cls, v):
      # 简单检查路径合法性,实际可更复杂
      if not v.strip():
      raise ValueError('文件路径不能为空')
      return v

      def write_file(params: WriteFileParams):
      try:
      with open(params.file_path, 'w', encoding='utf – 8') as f:
      f.write(params.content)
      return "文件写入成功"
      except Exception as e:
      print(f"文件写入失败: {e}")
      return None

    • 工具描述添加:在工具选择逻辑中,添加文件写入工具的描述。

      tools = [
      {
      "name": "calculate",
      "parameters": {
      "type": "object",
      "properties": {
      "num1": {"type": "number"},
      "num2": {"type": "number"},
      "operation": {"type": "string"}
      },
      "required": ["num1", "num2", "operation"]
      },
      "description": "用于执行简单的数学运算"
      },
      {
      "name": "search",
      "parameters": {
      "type": "object",
      "properties": {
      "query": {"type": "string"},
      "api_key": {"type": "string"}
      },
      "required": ["query", "api_key"]
      },
      "description": "用于搜索信息"
      },
      {
      "name": "read_file",
      "parameters": {
      "type": "object",
      "properties": {
      "file_path": {"type": "string"}
      },
      "required": ["file_path"]
      },
      "description": "用于读取本地文件内容"
      },
      {
      "name": "write_file",
      "parameters": {
      "type": "object",
      "properties": {
      "file_path": {"type": "string"},
      "content": {"type": "string"}
      },
      "required": ["file_path", "content"]
      },
      "description": "用于将内容写入本地文件"
      }
      ]

    • 结果解析与回答生成调整:添加文件写入工具的结果解析函数,并在 react_agent 函数中处理文件写入工具的调用和结果解析。

      def parse_write_file_result(result):
      if result:
      return result
      return "文件写入失败"

      def react_agent(user_input, api_key):
      tool_name, tool_params = choose_tool(user_input,api_key):
      if tool_name == 'calculate':
      params = CalculatorParams(**tool_params)
      result = calculate(params)
      parsed_result = parse_calculator_result(result)
      elif tool_name =='search':
      params = SearchParams(**tool_params)
      result = search(params)
      parsed_result = parse_search_result(result)
      elif tool_name =='read_file':
      params = ReadFileParams(**tool_params)
      result = read_file(params)
      parsed_result = parse_read_file_result(result)
      elif tool_name == 'write_file':
      params = WriteFileParams(**tool_params)
      result = write_file(params)
      parsed_result = parse_write_file_result(result)
      else:
      return "无法确定要使用的工具"
      return generate_answer(parsed_result, api_key)

    • 优化后的完整代码:

  • import requests
    import json
    import os
    from pydantic import BaseModel, validator

    # 计算器工具
    class CalculatorParams(BaseModel):
    num1: float
    num2: float
    operation: str

    @validator('operation')
    def valid_operation(cls, v):
    valid_ops = ['+', '-', '*', '/']
    if v not in valid_ops:
    raise ValueError('不支持的运算操作')
    return v

    def calculate(params: CalculatorParams):
    if params.operation == '+':
    return params.num1 + params.num2
    elif params.operation == '-':
    return params.num1 – params.num2
    elif params.operation == '*':
    return params.num1 * params.num2
    elif params.operation == '/':
    if params.num2 == 0:
    print("除数不能为零")
    return None
    return params.num1 / params.num2

    # 搜索工具
    class SearchParams(BaseModel):
    query: str
    api_key: str

    @validator('query')
    def query_not_empty(cls, v):
    if not v.strip():
    raise ValueError('查询内容不能为空')
    return v

    @validator('api_key')
    def api_key_not_empty(cls, v):
    if not v.strip():
    raise ValueError('api_key不能为空')
    return v

    def search(params: SearchParams):
    base_url = "https://api.example.com/search"
    params_dict = {
    "key": params.api_key,
    "q": params.query
    }
    try:
    response = requests.get(base_url, params=params_dict)
    response.raise_for_status()
    return response.json()
    except requests.RequestException as e:
    print(f"搜索失败: {e}")
    return None

    # 文件读取工具
    class ReadFileParams(BaseModel):
    file_path: str

    @validator('file_path')
    def file_path_exists(cls, v):
    if not os.path.exists(v):
    raise ValueError('文件路径不存在')
    return v

    def read_file(params: ReadFileParams):
    try:
    with open(params.file_path, 'r', encoding='utf – 8') as f:
    return f.read()
    except Exception as e:
    print(f"读取文件失败: {e}")
    return None

    # 文件写入工具
    class WriteFileParams(BaseModel):
    file_path: str
    content: str

    @validator('file_path')
    def file_path_valid(cls, v):
    if not v.strip():
    raise ValueError('文件路径不能为空')
    return v

    def write_file(params: WriteFileParams):
    try:
    with open(params.file_path, 'w', encoding='utf – 8') as f:
    f.write(params.content)
    return "文件写入成功"
    except Exception as e:
    print(f"文件写入失败: {e}")
    return None

    # 模拟通义千问API调用(实际需替换为真实API调用)
    def call_qwen_api(api_key, prompt):
    url = "https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions"
    headers = {
    "Content-Type": "application/json",
    "Authorization": f"Bearer {api_key}"
    }
    data = {
    "model": "qwen-plus",
    "messages": [
    {"role": "user", "content": prompt}
    ],
    "stream": False
    }
    try:
    response = requests.post(url, headers=headers, json=data)
    response.raise_for_status()
    return response.json()
    except requests.RequestException as e:
    print(f"调用通义千问API失败: {e}")
    return None

    def choose_tool(user_input, api_key):
    tools = [
    {
    "name": "calculate",
    "parameters": {
    "type": "object",
    "properties": {
    "num1": {"type": "number"},
    "num2": {"type": "number"},
    "operation": {"type": "string"}
    },
    "required": ["num1", "num2", "operation"]
    },
    "description": "用于执行简单的数学运算"
    },
    {
    "name": "search",
    "parameters": {
    "type": "object",
    "properties": {
    "query": {"type": "string"},
    "api_key": {"type": "string"}
    },
    "required": ["query", "api_key"]
    },
    "description": "用于搜索信息"
    },
    {
    "name": "read_file",
    "parameters": {
    "type": "object",
    "properties": {
    "file_path": {"type": "string"}
    },
    "required": ["file_path"]
    },
    "description": "用于读取本地文件内容"
    },
    {
    "name": "write_file",
    "parameters": {
    "type": "object",
    "properties": {
    "file_path": {"type": "string"},
    "content": {"type": "string"}
    },
    "required": ["file_path", "content"]
    },
    "description": "用于将内容写入本地文件"
    }
    ]
    prompt = {
    "user_input": user_input,
    "functions": tools
    }
    response = call_qwen_api(api_key, json.dumps(prompt))
    if "function_call" in response:
    function_call = response["function_call"]
    return function_call["name"], function_call["parameters"]
    return None, None

    def parse_calculator_result(result):
    if result is not None:
    return f"计算结果为{result}"
    return ""

    def parse_search_result(result):
    if result:
    return f"搜索结果: {result.get('results', [])}"
    return ""

    def parse_read_file_result(result):
    if result:
    return f"文件内容: {result}"
    return ""

    def parse_write_file_result(result):
    if result:
    return result
    return "文件写入失败"

    def generate_answer(parsed_result, api_key):
    prompt = f"根据以下信息生成回答: {parsed_result}"
    response = call_qwen_api(api_key, prompt)
    if response:
    return response["choices"][0]["message"]["content"]
    return "无法生成回答"

    def react_agent(user_input, api_key):
    tool_name, tool_params = choose_tool(user_input, api_key)
    if tool_name == 'calculate':
    params = CalculatorParams(**tool_params)
    result = calculate(params)
    parsed_result = parse_calculator_result(result)
    elif tool_name =='search':
    params = SearchParams(**tool_params)
    result = search(params)
    parsed_result = parse_search_result(result)
    elif tool_name =='read_file':
    params = ReadFileParams(**tool_params)
    result = read_file(params)
    parsed_result = parse_read_file_result(result)
    elif tool_name == 'write_file':
    params = WriteFileParams(**tool_params)
    result = write_file(params)
    parsed_result = parse_write_file_result(result)
    else:
    return "无法确定要使用的工具"
    return generate_answer(parsed_result, api_key)

    if __name__ == "__main__":
    api_key = "your_api_key"
    user_input = input("请输入你的问题: ")
    print(f"用户输入: {user_input}")
    tool_name, tool_params = choose_tool(user_input, api_key)
    print(f"选择的工具: {tool_name}")
    print(f"工具参数: {tool_params}")
    if tool_name == 'calculate':
    params = CalculatorParams(**tool_params)
    result = calculate(params)
    print(f"计算结果: {result}")
    parsed_result = parse_calculator_result(result)
    elif tool_name =='search':
    params = SearchParams(**tool_params)
    result = search(params)
    print(f"搜索结果: {result}")
    parsed_result = parse_search_result(result)
    elif tool_name =='read_file':
    params = ReadFileParams(**tool_params)
    result = read_file(params)
    print(f"文件读取结果: {result}")
    parsed_result = parse_read_file_result(result)
    elif tool_name == 'write_file':
    params = WriteFileParams(**tool_params)
    result = write_file(params)
    print(f"文件写入结果: {result}")
    parsed_result = parse_write_file_result(result)
    else:
    print("无法确定要使用的工具")
    exit()
    answer = generate_answer(parsed_result, api_key)
    print(f"生成的回答: {answer}")

    总结

    通过本周的学习,我们成功开发并优化了一个基于通义千问大模型的 ReAct Agent。从最初整合 Function Call、ReAct 范式等知识开发出支持多种工具的 Agent,到对其进行调试优化,包括修复工具选择和结果解析的问题,并增加新的文件写入工具,逐步提升了 Agent 的功能和稳定性。

    在实际应用中,可根据具体需求进一步扩展工具种类,优化工具选择逻辑以及结果解析方式,以更好地满足用户多样化的需求。同时,记得将本周知识点笔记以及优化后的 Agent 代码整理保存,方便后续参考和改进。如果在学习过程中遇到问题,可回顾本周所学内容,或者查阅相关文档进行解决。

    赞(0)
    未经允许不得转载:171主机测评 » Day13:迈向智能交互:从 ReAct Agent 开发到优化的全历程
    分享到: 更多 (0)

    评论 抢沙发

    • 昵称 (必填)
    • 邮箱 (必填)
    • 网址