Parse field description in function_tool.FunctionTool

[lienchen0526]

I'm currently using function_tool.FunctionTool to wrap my function in ADK with GOOGLE_AI. For example:

@function_tool.FunctionTool
def get_mailing_inbox(user_metadata: UserMetaData) -> Mails:
  ...

Both UserMetaData and Mails are a pydantic BaseModel. I gave the description to each fields clearly, hoped to make it visible to Gemini. However, I found the code here will only parse the description of parameter, but not return type (i.e. Mails) when we are using VERTEX_AI:

def from_function_with_options(
    func: Callable,
    variant: Literal['GOOGLE_AI', 'VERTEX_AI', 'DEFAULT'] = 'GOOGLE_AI',
) -> 'types.FunctionDeclaration':
    ...
    if variant == 'VERTEX_AI':
      declaration.parameters.required = (
          function_parameter_parse_util._get_required_fields(
              declaration.parameters
          )
      )
    if not variant == 'VERTEX_AI':
      return declaration
  ...

I wonder if we can make the description of the returned type also be parsed even we are using GOOGLE_AI? Or there are some limitations I missed? Many thank.

If it is feasible, I have a pull request: #88

Reference:

• https://ai.google.dev/api/caching#FunctionDeclaration

Best Regards,
Jerry

My LDAP: jerrylien@google.com

[Jacksunwei]

GOOGLE_AI backend will raise error if response schema is set. Only Vertex AI backend takes response.

In the meanwhile, we want to converge to using genai's conversion function. We just haven't gotten bandwidth to test whether it's equivalent to our current implementation. If you're interest, feel free to send us CLs. We use piper as SOT.

https://github.com/googleapis/python-genai/blob/main/google/genai/types.py#L1332

[boyangsvl]

BTW I don't think it works by annotating with @function_tool.FunctionTool. You don't have to use any annotation for a function to be used as a tool.