strict tool schema enforcement

[richardhundt]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

The OpenAI API allows non-strict tool schemas, but in validate_input_tools you just raise an exception if strict is not true. I've ended up writing my own client using httpx but I'd prefer to not have to maintain it, given that this just seems like an oversight. You're adding constraints which aren't present in the API.

To Reproduce

Just pass any non-strict tool schema.

Code snippets

OS

macOS

Python version

Python 3.10

Library version

all versions

[RobertCraigie]

We only validate tools are strict in the .parse() method, as that's the entire point of the method, if you want to use non strict tools you should use .create()

[richardhundt]

No you don't. You validate them somewhere down the stream call stack. I'm not calling parse.

EDIT: yep, it's right at the top of stream. But okay, I'll use create. Thanks for the tip.

[richardhundt]

openai-python/src/openai/resources/chat/completions/completions.py

Line 2879 in 4756247

_validate_input_tools(tools)

[richardhundt]

Another thing you might want to look at is the documentation for .parse()

        This method will also automatically parse `function` tool calls if:
        - You use the `openai.pydantic_function_tool()` helper method
        - You mark your tool schema with `"strict": True`

That reads as if you can opt out by not marking the tool schema with "strict": True

[richardhundt]

and .parse() would probably also break under these conditions:

>>> class Foo(BaseModel):
...   model_config=ConfigDict(extra="allow")
...   x: float
... 
>>> foo = Foo(x=4.2, y=6.9)
>>> foo.x
4.2
>>> foo.y
6.9
>>> Foo.model_json_schema()
{'additionalProperties': True, 'properties': {'x': {'title': 'X', 'type': 'number'}}, 'required': ['x'], 'title': 'Foo', 'type': 'object'}

[nomiveritas]

The .parse() method used with Pydantic models can break if the model allows extra fields (extra="allow").

Example:

class Foo(BaseModel):
model_config = ConfigDict(extra="allow")
x: float

foo = Foo(x=4.2, y=6.9) # 'y' is an extra field

In this case, .parse() does not handle extra fields properly and may cause an error.

---

Quick fix

1. If extra fields are not needed, disable extra="allow":

class Foo(BaseModel):
model_config = ConfigDict(extra="forbid") # or default
x: float

2. If extra fields are needed, use the .model_validate() method instead of .parse():

foo = Foo.model_validate({'x': 4.2, 'y': 6.9})
print(foo.x) # 4.2
print(foo.y) # 6.9

This way, the model won’t break, and the extra fields are preserved.