在语义、数学、推理、代码、知识等不同角度的数据集上测评显示,ChatGLM3-6B-Base 具有在10B以下的基础模型中最强的性能。ChatGLM3-6B采用了全新设计的Prompt格式,除正常的多轮对话外。同时原生支持工具调用(Function Call)、代码执行(Code Interpreter)和Agent任务等复杂场景。本文主要通过天气查询例子介绍了在tool_registry.py中注册新的工具来增强模型能力。
可以直接调用LangChain自带的工具(比如,ArXiv),也可以调用自定义的工具。LangChain自带的部分工具[2],如下所示:
一.自定义天气查询工具
1.Weather类
可以参考Tool/Weather.py以及Tool/Weather.yaml文件,继承BaseTool类,重载_run()方法,如下所示:
class Weather(BaseTool): # 天气查询工具
name = "weather"
description = "Use for searching weather at a specific location"
def __init__(self):
super().__init__()
def get_weather(self, location):
api_key = os.environ["SENIVERSE_KEY"]
url = f"https://api.seniverse.com/v3/weather/now.json?key={api_key}&location={location}&language=zh-Hans&unit=c"
response = requests.get(url)
if response.status_code == 200:
data = response.json()
weather = {
"temperature": data["results"][0]["now"]["temperature"],
"description": data["results"][0]["now"]["text"],
}
return weather
else:
raise Exception(
f"Failed to retrieve weather: {response.status_code}")
def _run(self, para: str) -> str:
return self.get_weather(para)
2.weather.yaml文件
weather.yaml文件内容,如下所示:
name: weather
description: Search the current weather of a city
parameters:
type: object
properties:
city:
type: string
description: City name
required:
- city
二.自定义天气查询工具调用
自定义天气查询工具调用,在main.py中导入Weather工具。如下所示:
run_tool([Weather()], llm, [
"今天北京天气怎么样?",
"What's the weather like in Shanghai today",
])
其中,run_tool()函数实现如下所示:
def run_tool(tools, llm, prompt_chain: List[str]):
loaded_tolls = [] # 用于存储加载的工具
for tool in tools: # 逐个加载工具
if isinstance(tool, str):
loaded_tolls.append(load_tools([tool], llm=llm)[0]) # load_tools返回的是一个列表
else:
loaded_tolls.append(tool) # 如果是自定义的工具,直接添加到列表中
agent = initialize_agent( # 初始化agent
loaded_tolls, llm,
agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION, # agent类型:使用结构化聊天的agent
verbose=True,
handle_parsing_errors=True
)
for prompt in prompt_chain: # 逐个输入prompt
agent.run(prompt)
1.load_tools()函数
根据工具名字加载相应的工具,如下所示:
def load_tools(
tool_names: List[str],
llm: Optional[BaseLanguageModel] = None,
callbacks: Callbacks = None,
**kwargs: Any,
) -> List[BaseTool]:
2.initialize_agent()函数
根据工具列表和LLM加载一个agent executor,如下所示:
def initialize_agent(
tools: Sequence[BaseTool],
llm: BaseLanguageModel,
agent: Optional[AgentType] = None,
callback_manager: Optional[BaseCallbackManager] = None,
agent_path: Optional[str] = None,
agent_kwargs: Optional[dict] = None,
*,
tags: Optional[Sequence[str]] = None,
**kwargs: Any,
) -> AgentExecutor:
其中,agent默认为AgentType.ZERO_SHOT_REACT_DESCRIPTION
。本文中使用为AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION
,一种为聊天模型优化的zero-shot react agent,该agent能够调用具有多个输入的工具。
3.run()函数
执行链的便捷方法,这个方法与Chain.__call__
之间的主要区别在于,这个方法期望将输入直接作为位置参数或关键字参数传递,而Chain.__call__
期望一个包含所有输入的单一输入字典。如下所示:
def run(
self,
*args: Any,
callbacks: Callbacks = None,
tags: Optional[List[str]] = None,
metadata: Optional[Dict[str, Any]] = None,
**kwargs: Any,
) -> Any:
4.结果分析
结果输出,如下所示:
> Entering new AgentExecutor chain...
======
======
Action:
``
{"action": "weather", "action_input": "北京"}
``
Observation: {'temperature': '20', 'description': '晴'}
Thought:======
======
Action:
``
{"action": "Final Answer", "action_input": "根据查询结果,北京今天的天气是晴,气温为20℃。"}
``
> Finished chain.
> Entering new AgentExecutor chain...
======
======
Action:
``
{"action": "weather", "action_input": "Shanghai"}
``
Observation: {'temperature': '20', 'description': '晴'}
Thought:======
======
Action:
``
{"action": "Final Answer", "action_input": "根据最新的天气数据,今天上海的天气情况是晴朗的,气温为20℃。"}
``
> Finished chain.
刚开始的时候没有找到识别实体city的地方,后面调试ChatGLM3/langchain_demo/ChatGLM3.py->_call()
时发现了一个巨长的prompt,这不就是zero-prompt(AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION)吗?顺便吐槽下LangChain的代码真的不好调试。
三.注册工具增强LLM能力
1.注册工具
可以通过在tool_registry.py中注册新的工具来增强模型的能力。只需要使用@register_tool装饰函数即可完成注册。对于工具声明,函数名称即为工具的名称,函数docstring即为工具的说明;对于工具的参数,使用Annotated[typ: type, description: str, required: bool]标注参数的类型、描述和是否必须。将get_weather()函数进行注册,如下所示:
@register_tool
def get_weather( # 工具函数
city_name: Annotated[str, 'The name of the city to be queried', True],
) -> str:
"""
Get the current weather for `city_name`
"""
if not isinstance(city_name, str): # 参数类型检查
raise TypeError("City name must be a string")
key_selection = { # 选择的键
"current_condition": ["temp_C", "FeelsLikeC", "humidity", "weatherDesc", "observation_time"],
}
import requests
try:
resp = requests.get(f"https://wttr.in/{city_name}?format=j1")
resp.raise_for_status()
resp = resp.json()
ret = {k: {_v: resp[k][0][_v] for _v in v} for k, v in key_selection.items()}
except:
import traceback
ret = "Error encountered while fetching weather data!\n" + traceback.format_exc()
return str(ret)
具体工具注册实现方式@register_tool装饰函数,如下所示:
def register_tool(func: callable): # 注册工具
tool_name = func.__name__ # 工具名
tool_description = inspect.getdoc(func).strip() # 工具描述
python_params = inspect.signature(func).parameters # 工具参数
tool_params = [] # 工具参数描述
for name, param in python_params.items(): # 遍历参数
annotation = param.annotation # 参数注解
if annotation is inspect.Parameter.empty:
raise TypeError(f"Parameter `{name}` missing type annotation") # 参数缺少注解
if get_origin(annotation) != Annotated: # 参数注解不是Annotated
raise TypeError(f"Annotation type for `{name}` must be typing.Annotated") # 参数注解必须是Annotated
typ, (description, required) = annotation.__origin__, annotation.__metadata__ # 参数类型, 参数描述, 是否必须
typ: str = str(typ) if isinstance(typ, GenericAlias) else typ.__name__ # 参数类型名
if not isinstance(description, str): # 参数描述必须是字符串
raise TypeError(f"Description for `{name}` must be a string")
if not isinstance(required, bool): # 是否必须必须是布尔值
raise TypeError(f"Required for `{name}` must be a bool")
tool_params.append({ # 添加参数描述
"name": name,
"description": description,
"type": typ,
"required": required
})
tool_def = { # 工具定义
"name": tool_name,
"description": tool_description,
"params": tool_params
}
print("[registered tool] " + pformat(tool_def)) # 打印工具定义
_TOOL_HOOKS[tool_name] = func # 注册工具
_TOOL_DESCRIPTIONS[tool_name] = tool_def # 添加工具定义
return func
2.调用工具
参考文件ChatGLM3/tool_using/openai_api_demo.py
,如下所示:
def main():
messages = [ # 对话信息
system_info,
{
"role": "user",
"content": "帮我查询北京的天气怎么样",
}
]
response = openai.ChatCompletion.create( # 调用OpenAI API
model="chatglm3",
messages=messages,
temperature=0,
return_function_call=True
)
function_call = json.loads(response.choices[0].message.content) # 获取函数调用信息
logger.info(f"Function Call Response: {function_call}") # 打印函数调用信息
tool_response = dispatch_tool(function_call["name"], function_call["parameters"]) # 调用函数
logger.info(f"Tool Call Response: {tool_response}") # 打印函数调用结果
messages = response.choices[0].history # 获取历史对话信息
messages.append(
{
"role": "observation",
"content": tool_response, # 调用函数返回结果
}
)
response = openai.ChatCompletion.create( # 调用OpenAI API
model="chatglm3",
messages=messages,
temperature=0,
)
logger.info(response.choices[0].message.content) # 打印对话结果
参考文献: