with_structured_output format depends on whether we're using pydantic proper or pydantic.v1 namespace

[eyurtsev]

Privileged issue

• I am a LangChain maintainer, or was asked directly by a LangChain maintainer to create an issue here.

Issue Content

from pydantic.v1 import BaseModel
from pydantic import BaseModel as BaseModelV2

class Answer(BaseModel):
    answer: str


class Answer2(BaseModelV2):
    """"The answer."""
    answer: str


from langchain_openai import ChatOpenAI

model = ChatOpenAI()

model.with_structured_output(Answer).invoke('the answer is foo') # <-- Returns pydantic object

model.with_structured_output(Answer2).invoke('the answer is foo') # <--- Returns dict

[eyurtsev]

There's a bug in the chat model when checking whether a pydantic model (from pydantic import BaseModel) ends up being classified as not a pydantic model.

But if we fix that logic, we get stuck here:

langchain/libs/core/langchain_core/output_parsers/openai_tools.py

Line 177 in 232908a

tools: List[Type[BaseModel]]

The code

class PydanticToolsParser(JsonOutputToolsParser):
    """Parse tools from OpenAI response."""

    tools: List[Type[BaseModel]]

Which will not work since JsonOutputToolsParser inherits from BaseModel as well

[hinthornw]

It also emits the wrong schema in general for nested objects