Function calling mode `ANY` + function declaration without parameters = `INVALID_ARGUMENT`

[DouweM]

Using FunctionCallingConfigMode.ANY with a FunctionDeclarationDict without parameters results in this error:

google.genai.errors.ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Request contains an invalid argument.', 'status': 'INVALID_ARGUMENT'}}

Code to reproduce:

import os

from google import genai
from google.genai.types import (
    FunctionCallingConfigDict,
    FunctionCallingConfigMode,
    FunctionDeclarationDict,
    GenerateContentConfigDict,
    ToolConfigDict,
    ToolDict,
)

client = genai.Client(api_key=os.getenv('GEMINI_API_KEY'))

result = client.models.generate_content(
    model='gemini-2.0-flash',
    contents='Test 123',
    config=GenerateContentConfigDict(
        tools=[
            ToolDict(function_declarations=[FunctionDeclarationDict(name='something')]),
        ],
        tool_config=ToolConfigDict(
            function_calling_config=FunctionCallingConfigDict(
                mode=FunctionCallingConfigMode.ANY,
            ),
        ),
    ),
)

I'm a maintainer of Pydantic AI, this was reported to us by a few users at pydantic/pydantic-ai#1990. One of them says it used to work last week but does no longer, so this look like an API-side regression. Please let me know if there's a more appropriate place to file API issues.

[DouweM]

It seems to work when we set parameters to an object schema without any properties:

import os

from google import genai
from google.genai.types import (
    FunctionCallingConfigDict,
    FunctionCallingConfigMode,
    FunctionDeclarationDict,
    GenerateContentConfigDict,
    SchemaDict,
    ToolConfigDict,
    ToolDict,
    Type,
)

client = genai.Client(api_key=os.getenv('GEMINI_API_KEY'))

result = client.models.generate_content(
    model='gemini-2.0-flash',
    contents='Test 123',
    config=GenerateContentConfigDict(
        tools=[
            ToolDict(
                function_declarations=[
                    FunctionDeclarationDict(
                        name='something',
                        parameters=SchemaDict(type=Type.OBJECT),
                    )
                ]
            ),
        ],
        tool_config=ToolConfigDict(
            function_calling_config=FunctionCallingConfigDict(
                mode=FunctionCallingConfigMode.ANY,
            ),
        ),
    ),
)

print(result)

[ewsiegel]

Seconding this issue -- almost this exact execution worked for me on Friday 6/13 around 5pm EDT, but on Sunday 6/15 around 1pm EDT it was no longer working.

[wanlin31]

Hello @DouweM and @ewsiegel ,

Thank you so much for surface this issue.

I have tried the exact code with the current sdk and sdk synced back to (#964), I got same error response for both. Seems like not an sdk issue, will report internally to investigate.

The error message seems like pointing out the parameter's type must be object:

"Fail to execute model for flow_id: gemini-2.0-flash-001-flow_runner_text_eu_TEXT_32k_1m\\nError: Schema.type must be OBJECT"

Your workaround seems work, in the mean time, can we switch to that?

Thank you

[ewsiegel]

This workaround worked for me, thanks!