Error if function_tool has a Mapping input

[TimoVink]

Describe the bug

If an agent is provided with a function_tool that has a Mapping or dict as its input, an HTTP 400 invalid_function_parameters error is returned when the agent runs.

Example

@function_tool
def foo(x: Mapping[str,str]) -> str:
    return 'bar'

foo_agent = Agent(name='Foo Agent', tools=[foo])
await Runner.run(foo_agent, input='foo')

Error:

Error getting response: Error code: 400
{
    'error': {
        'message': "Invalid schema for function 'foo': In context=(), 'required' is required to be supplied and to be an array including every key in properties. Extra required key 'x' supplied.", 
        'type': 'invalid_request_error',
        'param': 'tools[0].parameters',
        'code': 'invalid_function_parameters'
    }
}
(request_id: req_f3213fe15078c00d899a910c5567fed4)

If I check foo.params_json_schema, I get:

{'properties': {'x': {'additionalProperties': {'type': 'string'},
   'title': 'X',
   'type': 'object'}},
 'required': ['x'],
 'title': 'foo_args',
 'type': 'object',
 'additionalProperties': False}

Which looks valid to me at a glance, but I'm not very familiar with the JSON Schema spec.

Debug information

• Agents SDK: v0.0.6
• Python: v3.11.9
• Pydantic: v2.10.6

[DanieleMorotti]

Hi, in short, you can fix the error by creating a Pydantic object such as:

from pydantic import BaseModel

class MyObject(BaseModel):
    parameter_name: str

and then you declare it as the type of the parameter:

@function_tool
def foo(x: MyObject) -> str:
...

Because there are several issues regarding support for generic dict or Mapping types instead of objects.

[TimoVink]

Hi, thanks for the response.

My original goal was to model something easily coercible to a pd.Series, but in a way that gives more type hinting than just using something like list[list[Any]]

With pd.Series itself not working, nor Tuple (#302) nor Mapping, I am now doing something like this:

class DataPoint(BaseModel):
    x: str
    y: float

class DataSeries(BaseModel):
    title: str
    data: list[DataPoint]

... which works but feels more verbose than necessary. Though perhaps this is actually clearer for the model than anything else I could do.

Do you know where the limitation for Mapping is coming from? Is it due to limitations in JSON Schema? Pydantic? This library? Then OpenAI API? Happy to close or move this issue to the best place if needed!

[rm-openai]

Looking into it - thanks for the report

[rm-openai]

Looks like the issue is that a schema like this is not valid in strict mode:

class Foo(BaseModel):
  bar: Mapping[str, int]

specifically because of the 'additionalProperties': {'type': 'string'} in the generated schema. Basically for structured outputs, we can guarantee correct JSON but only for a constrained JSON schema. This one isn't constrained because it's not clear how to fill in bar in a valid way - when would you stop adding keys, for example?

We should actually be raising an error at runtime for this. I will fix that, but for now you can use @function_tool(strict=False). Or alternatively, something more typed e.g.

class MyData(BaseModel):
  value: str
  some_other_value: int
  ...

@function_tool
def foo(data: MyData)

[DanieleMorotti]

Basically for structured outputs, we can guarantee correct JSON but only for a constrained JSON schema.

Ok, in fact, I thought this was the reason for the error. Thanks for the clarification @rm-openai

[rm-openai]

Closing this out as resolved with the assert in #356. Feel free to reopen/new issue for followups!

[johnandersen777]

@rm-openai #356 broke otherwise working json_schema output code for me :(

For example: It's useful to be able to generate GitHub Actions style workflows where with is used for per action arguments. Of course we could generate potential actions and then create pydantic classes for them, but that's a lot for what was otherwise working.

Is there something we could do to keep that is True? Or am I SOL?