大模型从入门到应用——LangChain：代理（Agents）-[工具（Tools）：人工确认工具验证和Tools作为OpenAI函数]
分类目录：《大模型从入门到应用》总目录
LangChain系列文章：
- 基础知识
- 快速入门
- 安装与环境配置
- 链（Chains）、代理（Agent:）和记忆（Memory）
- 快速开发聊天模型
- 模型（Models）
- 基础知识
- 大型语言模型（LLMs）
- 基础知识
- LLM的异步API、自定义LLM包装器、虚假LLM和人类输入LLM（Human Input LLM）
- 缓存LLM的调用结果
- 加载与保存LLM类、流式传输LLM与Chat Model响应和跟踪tokens使用情况
- 聊天模型（Chat Models）
- 基础知识
- 使用少量示例和响应流式传输
- 文本嵌入模型
- Aleph Alpha、Amazon Bedrock、Azure OpenAI、Cohere等
- Embaas、Fake Embeddings、Google Vertex AI PaLM等
- 提示（Prompts）
- 基础知识
- 提示模板
- 基础知识
- 连接到特征存储
- 创建自定义提示模板和含有Few-Shot示例的提示模板
- 部分填充的提示模板和提示合成
- 序列化提示信息
- 示例选择器（Example Selectors）
- 输出解析器（Output Parsers）
- 记忆（Memory）
- 基础知识
- 记忆的类型
- 会话缓存记忆、会话缓存窗口记忆和实体记忆
- 对话知识图谱记忆、对话摘要记忆和会话摘要缓冲记忆
- 对话令牌缓冲存储器和基于向量存储的记忆
- 将记忆添加到LangChain组件中
- 自定义对话记忆与自定义记忆类
- 聊天消息记录
- 记忆的存储与应用
- 索引（Indexes）
- 基础知识
- 文档加载器（Document Loaders）
- 文本分割器（Text Splitters）
- 向量存储器（Vectorstores）
- 检索器（Retrievers）
- 链（Chains）
- 基础知识
- 通用功能
- 自定义Chain和Chain的异步API
- LLMChain和RouterChain
- SequentialChain和TransformationChain
- 链的保存（序列化）与加载（反序列化）
- 链与索引
- 文档分析和基于文档的聊天
- 问答的基础知识
- 图问答（Graph QA）和带来源的问答（Q&A with Sources）
- 检索式问答
- 文本摘要（Summarization）、HyDE和向量数据库的文本生成
- 代理（Agents）
- 基础知识
- 代理类型
- 自定义代理（Custom Agent）
- 自定义MRKL代理
- 带有ChatModel的LLM聊天自定义代理和自定义多操作代理（Custom MultiAction Agent）
- 工具
- 基础知识
- 自定义工具（Custom Tools）
- 多输入工具和工具输入模式
- 人工确认工具验证和Tools作为OpenAI函数
- 工具包（Toolkit）
- 代理执行器（Agent Executor）
- 结合使用Agent和VectorStore
- 使用Agents的异步API和创建ChatGPT克隆
- 处理解析错误、访问中间步骤和限制最大迭代次数
- 为代理程序设置超时时间和限制最大迭代次数和为代理程序和其工具添加共享内存
- 计划与执行
- 回调函数（Callbacks）
人工确认工具验证
本节演示如何为任何工具添加人工确认验证，我们将使用HumanApprovalCallbackhandler完成此操作。假设我们需要使用ShellTool，将此工具添加到自动化流程中会带来明显的风险。我们将看看如何强制对输入到该工具的内容进行手动人工确认。我们通常建议不要使用ShellTool。它有很多被误用的方式，并且在大多数情况下并不需要使用它。我们这里只是为了演示目的才使用它。
from langchain.callbacks import HumanApprovalCallbackHandler
from langchain.tools import ShellTool
tool = ShellTool()
print(tool.run('echo Hello World!'))
输出：
Hello World!
添加人工确认
将默认的HumanApprovalCallbackHandler添加到工具中，这样在实际执行命令之前，用户必须手动批准工具的每个输入。
tool = ShellTool(callbacks=[HumanApprovalCallbackHandler()])
print(tool.run("ls /usr"))
日志输出与输入：
Do you approve of the following input? Anything except 'Y'/'Yes' (case-insensitive) will be treated as a no.
ls /usr
yes
X11
X11R6
bin
lib
libexec
local
sbin
share
standalone
输入：
print(tool.run("ls /private"))
日志输出与输入：
Do you approve of the following input? Anything except 'Y'/'Yes' (case-insensitive) will be treated as a no.
ls /private
no
---------------------------------------------------------------------------
HumanRejectedException Traceback (most recent call last)
Cell In[17], line 1
----> 1 print(tool.run("ls /private"))
File ~/langchain/langchain/tools/base.py:257, in BaseTool.run(self, tool_input, verbose, start_color, color, callbacks, **kwargs)
255 # TODO: maybe also pass through run_manager is _run supports kwargs
256 new_arg_supported = signature(self._run).parameters.get("run_manager")
--> 257 run_manager = callback_manager.on_tool_start(
258 {"name": self.name, "description": self.description},
259 tool_input if isinstance(tool_input, str) else str(tool_input),
260 color=start_color,
261 **kwargs,
262 )
263 try:
264 tool_args, tool_kwargs = self._to_args_and_kwargs(parsed_input)
File ~/langchain/langchain/callbacks/manager.py:672, in CallbackManager.on_tool_start(self, serialized, input_str, run_id, parent_run_id, **kwargs)
669 if run_id is None:
670 run_id = uuid4()
--> 672 _handle_event(
673 self.handlers,
674 "on_tool_start",
675 "ignore_agent",
676 serialized,
677 input_str,
678 run_id=run_id,
679 parent_run_id=self.parent_run_id,
680 **kwargs,
681 )
683 return CallbackManagerForToolRun(
684 run_id, self.handlers, self.inheritable_handlers, self.parent_run_id
685 )
File ~/langchain/langchain/callbacks/manager.py:157, in _handle_event(handlers, event_name, ignore_condition_name, *args, **kwargs)
155 except Exception as e:
156 if handler.raise_error:
--> 157 raise e
158 logging.warning(f"Error in {event_name} callback: {e}")
File ~/langchain/langchain/callbacks/manager.py:139, in _handle_event(handlers, event_name, ignore_condition_name, *args, **kwargs)
135 try:
136 if ignore_condition_name is None or not getattr(
137 handler, ignore_condition_name
138 ):
--> 139 getattr(handler, event_name)(*args, **kwargs)
140 except NotImplementedError as e:
141 if event_name == "on_chat_model_start":
File ~/langchain/langchain/callbacks/human.py:48, in HumanApprovalCallbackHandler.on_tool_start(self, serialized, input_str, run_id, parent_run_id, **kwargs)
38 def on_tool_start(
39 self,
40 serialized: Dict[str, Any],
(...)
45 **kwargs: Any,
46 ) -> Any:
47 if self._should_check(serialized) and not self._approve(input_str):
---> 48 raise HumanRejectedException(
49 f"Inputs {input_str} to tool {serialized} were rejected."
50 )
HumanRejectedException: Inputs ls /private to tool {'name': 'terminal', 'description': 'Run shell commands on this MacOS machine.'} were rejected.
配置人工确认
假设我们有一个代理程序，接收多个工具，并且我们只希望在某些工具和某些输入上触发人工确认请求。我们可以配置回调处理程序来实现这一点。
from langchain.agents import load_tools
from langchain.agents import initialize_agent
from langchain.agents import AgentType
from langchain.llms import OpenAI
def _should_check(serialized_obj: dict) -> bool:
# Only require approval on ShellTool.
return serialized_obj.get("name") == "terminal"
def _approve(_input: str) -> bool:
if _input == "echo 'Hello World'":
return True
msg = (
"Do you approve of the following input? "
"Anything except 'Y'/'Yes' (case-insensitive) will be treated as a no."
)
msg += "\n\n" + _input + "\n"
resp = input(msg)
return resp.lower() in ("yes", "y")
callbacks = [HumanApprovalCallbackHandler(should_check=_should_check, approve=_approve)]
llm = OpenAI(temperature=0)
tools = load_tools(["wikipedia", "llm-math", "terminal"], llm=llm)
agent = initialize_agent(
tools,
llm,
agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
)
agent.run("It's 2023 now. How many years ago did Konrad Adenauer become Chancellor of Germany.", callbacks=callbacks)
输出：
'Konrad Adenauer became Chancellor of Germany in 1949, 74 years ago.'
输入：
agent.run("print 'Hello World' in the terminal", callbacks=callbacks)
输出：
'Hello World'
输入：
agent.run("list all directories in /private", callbacks=callbacks)
日志输出与输入：
ls /private
no
---------------------------------------------------------------------------
HumanRejectedException Traceback (most recent call last)
Cell In[39], line 1
----> 1 agent.run("list all directories in /private", callbacks=callbacks)
File ~/langchain/langchain/chains/base.py:236, in Chain.run(self, callbacks, *args, **kwargs)
234 if len(args) != 1:
235 raise ValueError("`run` supports only one positional argument.")
--> 236 return self(args[0], callbacks=callbacks)[self.output_keys[0]]
238 if kwargs and not args:
239 return self(kwargs, callbacks=callbacks)[self.output_keys[0]]
File ~/langchain/langchain/chains/base.py:140, in Chain.__call__(self, inputs, return_only_outputs, callbacks)
138 except (KeyboardInterrupt, Exception) as e:
139 run_manager.on_chain_error(e)
--> 140 raise e
141 run_manager.on_chain_end(outputs)
142 return self.prep_outputs(inputs, outputs, return_only_outputs)
File ~/langchain/langchain/chains/base.py:134, in Chain.__call__(self, inputs, return_only_outputs, callbacks)
128 run_manager = callback_manager.on_chain_start(
129 {"name": self.__class__.__name__},
130 inputs,
131 )
132 try:
133 outputs = (
--> 134 self._call(inputs, run_manager=run_manager)
135 if new_arg_supported
136 else self._call(inputs)
137 )
138 except (KeyboardInterrupt, Exception) as e:
139 run_manager.on_chain_error(e)
File ~/langchain/langchain/agents/agent.py:953, in AgentExecutor._call(self, inputs, run_manager)
951 # We now enter the agent loop (until it returns something).
952 while self._should_continue(iterations, time_elapsed):
--> 953 next_step_output = self._take_next_step(
954 name_to_tool_map,
955 color_mapping,
956 inputs,
957 intermediate_steps,
958 run_manager=run_manager,
959 )
960 if isinstance(next_step_output, AgentFinish):
961 return self._return(
962 next_step_output, intermediate_steps, run_manager=run_manager
963 )
File ~/langchain/langchain/agents/agent.py:820, in AgentExecutor._take_next_step(self, name_to_tool_map, color_mapping, inputs, intermediate_steps, run_manager)
818 tool_run_kwargs["llm_prefix"] = ""
819 # We then call the tool on the tool input to get an observation
--> 820 observation = tool.run(
821 agent_action.tool_input,
822 verbose=self.verbose,
823 color=color,
824 callbacks=run_manager.get_child() if run_manager else None,
825 **tool_run_kwargs,
826 )
827 else:
828 tool_run_kwargs = self.agent.tool_run_logging_kwargs()
File ~/langchain/langchain/tools/base.py:257, in BaseTool.run(self, tool_input, verbose, start_color, color, callbacks, **kwargs)
255 # TODO: maybe also pass through run_manager is _run supports kwargs
256 new_arg_supported = signature(self._run).parameters.get("run_manager")
--> 257 run_manager = callback_manager.on_tool_start(
258 {"name": self.name, "description": self.description},
259 tool_input if isinstance(tool_input, str) else str(tool_input),
260 color=start_color,
261 **kwargs,
262 )
263 try:
264 tool_args, tool_kwargs = self._to_args_and_kwargs(parsed_input)
File ~/langchain/langchain/callbacks/manager.py:672, in CallbackManager.on_tool_start(self, serialized, input_str, run_id, parent_run_id, **kwargs)
669 if run_id is None:
670 run_id = uuid4()
--> 672 _handle_event(
673 self.handlers,
674 "on_tool_start",
675 "ignore_agent",
676 serialized,
677 input_str,
678 run_id=run_id,
679 parent_run_id=self.parent_run_id,
680 **kwargs,
681 )
683 return CallbackManagerForToolRun(
684 run_id, self.handlers, self.inheritable_handlers, self.parent_run_id
685 )
File ~/langchain/langchain/callbacks/manager.py:157, in _handle_event(handlers, event_name, ignore_condition_name, *args, **kwargs)
155 except Exception as e:
156 if handler.raise_error:
--> 157 raise e
158 logging.warning(f"Error in {event_name} callback: {e}")
File ~/langchain/langchain/callbacks/manager.py:139, in _handle_event(handlers, event_name, ignore_condition_name, *args, **kwargs)
135 try:
136 if ignore_condition_name is None or not getattr(
137 handler, ignore_condition_name
138 ):
--> 139 getattr(handler, event_name)(*args, **kwargs)
140 except NotImplementedError as e:
141 if event_name == "on_chat_model_start":
File ~/langchain/langchain/callbacks/human.py:48, in HumanApprovalCallbackHandler.on_tool_start(self, serialized, input_str, run_id, parent_run_id, **kwargs)
38 def on_tool_start(
39 self,
40 serialized: Dict[str, Any],
(...)
45 **kwargs: Any,
46 ) -> Any:
47 if self._should_check(serialized) and not self._approve(input_str):
---> 48 raise HumanRejectedException(
49 f"Inputs {input_str} to tool {serialized} were rejected."
50 )
HumanRejectedException: Inputs ls /private to tool {'name': 'terminal', 'description': 'Run shell commands on this MacOS machine.'} were rejected.
Tools作为OpenAI函数
这个笔记本将介绍如何将LangChain的工具作为OpenAI函数使用。
from langchain.chat_models import ChatOpenAI
from langchain.schema import HumanMessage
model = ChatOpenAI(model="gpt-3.5-turbo-0613")
from langchain.tools import MoveFileTool, format_tool_to_openai_function
tools = [MoveFileTool()]
functions = [format_tool_to_openai_function(t) for t in tools]
message = model.predict_messages([HumanMessage(content='move file foo to bar')], functions=functions)
message
输出：
AIMessage(content='', additional_kwargs={'function_call': {'name': 'move_file', 'arguments': '{\n "source_path": "foo",\n "destination_path": "bar"\n}'}}, example=False)
输入：
message.additional_kwargs['function_call']
输出：
{'name': 'move_file',
'arguments': '{\n "source_path": "foo",\n "destination_path": "bar"\n}'}
参考文献：
[1] LangChain官方网站：https://www.langchain.com/
[2] LangChain ️ 中文网，跟着LangChain一起学LLM/GPT开发：https://www.langchain.com.cn/
[3] LangChain中文网 - LangChain 是一个用于开发由语言模型驱动的应用程序的框架：http://www.cnlangchain.com/