为何Langchain的with_structured_output忽略Pydantic模型的可选参数？

问题背景 #

我在测试Langchain通过with_structured_output从LLM获取类JSON结构化输出的功能时，发现直接传入Pydantic模型和传入其JSON Schema的处理逻辑不一致，尤其是对Optional字段和默认值的处理，不符合预期。

测试代码 #

from typing import Optional
from pydantic import BaseModel, Field
from langchain_openai import ChatOpenAI

class Joke(BaseModel):
    """Joke to tell user."""

    setup: str = Field(description="The setup of the joke")
    punchline: str = Field(description="The punchline to the joke")
    rating: Optional[int] = Field(None, description="How funny the joke is, from 1 to 10")
    nonsense: Optional[str] = Field(None, description="Placeholder, always return None")

class Joke2(BaseModel):
    """Joke to tell user."""

    setup: str = Field(description="The setup of the joke")
    punchline: str = Field(description="The punchline to the joke")
    rating: int = Field(description="How funny the joke is, from 1 to 10")
    nonsense: str = Field(description="Placeholder, always return None")

llm = ChatOpenAI(temperature=0, model_name="gpt-3.5-turbo")

测试结果与预期差异 #

我原本预期：

• 测试1和测试2的输出一致（返回rating且nonsense为null）
• 测试3和测试4的输出一致（返回rating且nonsense为字符串'None'）

但实际结果只有测试2和4符合预期，测试1和3出现偏差：

测试1：直接传入Joke模型 #

# 预期返回`rating`且`nonsense`为null
llm.with_structured_output(Joke).invoke("Tell me a joke about cats")
# 实际输出
{'setup': 'Why was the cat sitting on the computer?',
 'punchline': 'To keep an eye on the mouse!'}

问题：缺失rating和nonsense字段

测试2：传入Joke的JSON Schema #

# 预期返回`rating`且`nonsense`为null
llm.with_structured_output(Joke.model_json_schema()).invoke("Tell me a joke about cats")
# 实际输出
{'setup': 'Why was the cat sitting on the computer?',
 'punchline': 'To keep an eye on the mouse!',
 'rating': 8,
 'nonsense': None}

符合预期

测试3：直接传入Joke2模型 #

# 预期返回`rating`且`nonsense`为字符串'None'
llm.with_structured_output(Joke2).invoke("Tell me a joke about cats")
# 实际输出
{'setup': 'Why was the cat sitting on the computer?',
 'punchline': 'To keep an eye on the mouse!',
 'rating': 5,
 'nonsense': 'This joke is purr-fectly hilarious!'}

问题：nonsense未返回字符串'None'，而是生成了无关内容

测试4：传入Joke2的JSON Schema #

# 预期返回`rating`且`nonsense`为字符串'None'
llm.with_structured_output(Joke2.model_json_schema()).invoke("Tell me a joke about cats")
# 实际输出
{'setup': 'Why was the cat sitting on the computer?',
 'punchline': 'To keep an eye on the mouse!',
 'rating': 8,
 'nonsense': 'None'}

符合预期

疑问 #

为什么直接传入Pydantic模型的测试1和3不符合预期，而传入JSON Schema的测试2和4却能达到预期效果？

内容的提问来源于stack exchange，提问作者Peter