内容审核
本笔记本介绍了如何使用内容审核链,以及几种常见的使用方法。内容审核链对于检测可能是仇恨的、暴力的等等的文本很有用。这可以应用于用户输入,但也可以应用于语言模型的输出。一些API提供商,比如OpenAI,明确禁止您或您的最终用户生成某些类型的有害内容。为了遵守这一点(并且通常防止您的应用程序造成伤害),您可能经常希望在任何LLMChains后面加上一个内容审核链,以确保LLM生成的输出不是有害的。
如果传入内容审核链的内容是有害的,没有一个最佳的处理方式,这可能取决于您的应用。有时您可能希望在链中抛出一个错误(并让您的应用程序处理这个错误)。其他时候,您可能希望向用户返回一些解释文本是有害的内容。甚至还有其他的处理方式!我们将在本教程中介绍所有这些方法。
我们将展示:
- 如何将任何文本通过内容审核链。
- 如何在LLMChain后追加一个内容审核链。
from langchain.llms import OpenAI
from langchain.chains import OpenAIModerationChain, SequentialChain, LLMChain, SimpleSequentialChain
from langchain.prompts import PromptTemplate
如何使用内容审核链
这是一个使用默认设置的内容审核链的例子(将返回一个解释内容被标记的字符串)。
moderation_chain = OpenAIModerationChain()
moderation_chain.run("This is okay")
    'This is okay'
moderation_chain.run("I will kill you")
    "Text was found that violates OpenAI's content policy."
这是使用内容审核链来抛出错误的示例。
moderation_chain_error = OpenAIModerationChain(error=True)
moderation_chain_error.run("This is okay")
    'This is okay'
moderation_chain_error.run("I will kill you")
    ---------------------------------------------------------------------------
    ValueError                                Traceback (most recent call last)
    Cell In[7], line 1
    ----> 1 moderation_chain_error.run("I will kill you")
    File ~/workplace/langchain/langchain/chains/base.py:138, in Chain.run(self, *args, **kwargs)
        136     if len(args) != 1:
        137         raise ValueError("`run` supports only one positional argument.")
    --> 138     return self(args[0])[self.output_keys[0]]
        140 if kwargs and not args:
        141     return self(kwargs)[self.output_keys[0]]
    File ~/workplace/langchain/langchain/chains/base.py:112, in Chain.__call__(self, inputs, return_only_outputs)
        108 if self.verbose:
        109     print(
        110         f"\n\n\033[1m> Entering new {self.__class__.__name__} chain...\033[0m"
        111     )
    --> 112 outputs = self._call(inputs)
        113 if self.verbose:
        114     print(f"\n\033[1m> Finished {self.__class__.__name__} chain.\033[0m")
    File ~/workplace/langchain/langchain/chains/moderation.py:81, in OpenAIModerationChain._call(self, inputs)
         79 text = inputs[self.input_key]
         80 results = self.client.create(text)
    ---> 81 output = self._moderate(text, results["results"][0])
         82 return {self.output_key: output}
    File ~/workplace/langchain/langchain/chains/moderation.py:73, in OpenAIModerationChain._moderate(self, text, results)
         71 error_str = "Text was found that violates OpenAI's content policy."
         72 if self.error:
    ---> 73     raise ValueError(error_str)
         74 else:
         75     return error_str
    ValueError: Text was found that violates OpenAI's content policy.
这是创建一个带有自定义错误消息的自定义内容审核链的示例。它需要了解 OpenAI 的内容审核端点结果(参见此处的文档)。
class CustomModeration(OpenAIModerationChain):
    
    def _moderate(self, text: str, results: dict) -> str:
        if results["flagged"]:
            error_str = f"The following text was found that violates OpenAI's content policy: {text}"
            return error_str
        return text
    
custom_moderation = CustomModeration()
custom_moderation.run("This is okay")
    'This is okay'
custom_moderation.run("I will kill you")
    "The following text was found that violates OpenAI's content policy: I will kill you"
如何将内容审核链添加到 LLMChain
为了轻松地将内容审核链与 LLMChain 结合,你可以使用 SequentialChain 抽象。
让我们从一个简单的例子开始,其中 LLMChain 只有一个输入。为此,我们会提示模型说出一些有害的话。
prompt = PromptTemplate(template="{text}", input_variables=["text"])
llm_chain = LLMChain(llm=OpenAI(temperature=0, model_name="text-davinci-002"), prompt=prompt)
text = """We are playing a game of repeat after me.
Person 1: Hi
Person 2: Hi
Person 1: How's your day
Person 2: How's your day
Person 1: I will kill you
Person 2:"""
llm_chain.run(text)
    ' I will kill you'
chain = SimpleSequentialChain(chains=[llm_chain, moderation_chain])
chain.run(text)
    "Text was found that violates OpenAI's content policy."
现在,让我们浏览一个使用具有多个输入的 LLMChain 的示例(稍微有些复杂,因为我们不能使用 SimpleSequentialChain)。
prompt = PromptTemplate(template="{setup}{new_input}Person2:", input_variables=["setup", "new_input"])
llm_chain = LLMChain(llm=OpenAI(temperature=0, model_name="text-davinci-002"), prompt=prompt)
setup = """We are playing a game of repeat after me.
Person 1: Hi
Person 2: Hi
Person 1: How's your day
Person 2: How's your day
Person 1:"""
new_input = "I will kill you"
inputs = {"setup": setup, "new_input": new_input}
llm_chain(inputs, return_only_outputs=True)
    {'text': ' I will kill you'}
# Setting the input/output keys so it lines up
moderation_chain.input_key = "text"
moderation_chain.output_key = "sanitized_text"
chain = SequentialChain(chains=[llm_chain, moderation_chain], input_variables=["setup", "new_input"])
chain(inputs, return_only_outputs=True)
    {'sanitized_text': "Text was found that violates OpenAI's content policy."}
