继承Langchain类时super().__init__及自定义属性添加异常问题

初始可运行代码 #

以下代码可以正常运行：

from langchain_openai import ChatOpenAI

class LLM:
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name  

class OpenAILLM(LLM, ChatOpenAI):
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name

添加自定义属性后报错 #

当在OpenAILLM的__init__中添加自定义属性var时，创建对象失败：

from langchain_openai import ChatOpenAI

class LLM:
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name  

class OpenAILLM(LLM, ChatOpenAI):
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name
        self.var = 'var'

报错信息：

ValueError: "OpenAILLM" object has no field "var"

尝试类变量后的问题 #

将var声明为LLM的类变量后，get_var方法可以正常工作，但set_var方法触发相同错误：

from langchain_openai import ChatOpenAI

class LLM():
    var = 1
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name  

class OpenAILLM(LLM, ChatOpenAI):
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name
    def get_var(self):
        return self.var
    def set_var(self, var):
        print('self var', self.var)
        self.var = var
        return True

方法解析顺序(MRO) #

OpenAILLM的MRO如下：

(__main__.OpenAILLM,
 __main__.LLM,
 langchain_openai.chat_models.base.ChatOpenAI,
 langchain_core.language_models.chat_models.BaseChatModel,
 langchain_core.language_models.base.BaseLanguageModel,
 langchain_core.runnables.base.RunnableSerializable,
 langchain_core.load.serializable.Serializable,
 pydantic.v1.main.BaseModel,
 pydantic.v1.utils.Representation,
 langchain_core.runnables.base.Runnable,
 typing.Generic,
 abc.ABC,
 object)

原因分析 #

从MRO可以看到，OpenAILLM间接继承了Pydantic v1的BaseModel。Pydantic v1的BaseModel默认会严格校验实例属性，不允许添加未在类中声明的字段。直接赋值self.var时，Pydantic会拦截并报错，因为var不是模型定义的字段。

解决方案 #

方案1：在子类中声明自定义字段 #

在OpenAILLM类中显式声明var字段，符合Pydantic的规范：

from langchain_openai import ChatOpenAI
from pydantic.v1 import Field

class LLM:
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name  

class OpenAILLM(LLM, ChatOpenAI):
    var: str = Field(default='var')  # 声明字段，可设置默认值
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name

方案2：允许额外字段 #

通过设置Extra.allow，让Pydantic允许添加未声明的字段：

from langchain_openai import ChatOpenAI
from pydantic.v1 import Extra

class LLM:
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name  

class OpenAILLM(LLM, ChatOpenAI):
    class Config:
        extra = Extra.allow  # 允许额外字段
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name
        self.var = 'var'

方案3：使用对象的__dict__赋值（不推荐） #

绕过Pydantic的属性校验，直接修改__dict__，但可能破坏Pydantic的模型一致性：

from langchain_openai import ChatOpenAI

class LLM:
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name  

class OpenAILLM(LLM, ChatOpenAI):
    def __init__(self, model_name):
        super().__init__(model_name=model_name)
        self.model_name = model_name
        self.__dict__['var'] = 'var'  # 直接修改__dict__

内容的提问来源于stack exchange，提问作者Yadnesh Salvi