Improving how function docstring gets converted to tool's jsonschema

[salman1993]

Enhancement

This is a clone of this issue: modelcontextprotocol/python-sdk#226

Is your feature request related to a problem? Please describe.
Tool descriptions are not parsed as expected from the function docstring when using FastMCP. This affects tool calling performance.

Currently, FastMCP does some function inspection to create the docstring here:

• https://github.com/jlowin/fastmcp/blob/26969e80c3302934e05f5853ea8836607a3f557b/src/fastmcp/tools/tool.py#L235-L295

From my understanding, it creates a FuncMetadata model in pydantic which then gets converted to jsonschema.

Current behaviour:
If we have a tool such as:

def add_numbers(a: float, b: float) -> float:
    """
    Adds two numbers and returns the result.

    Args:
        a (float): The first number.
        b (float): The second number.

    Returns:
        float: The sum of a and b.
    """
    return a + b

it gets parsed into:

>>> from fastmcp.tools.tool import ParsedFunction
>>> print( ParsedFunction.from_function(fn=add_numbers) )
ParsedFunction(
    fn=<function add_numbers at 0x107ee5d00>,
    name='add_numbers',
    description='Adds two numbers and returns the result.\n\nArgs:\n    a (float): The first number.\n    b (float): The second number.\n\nReturns:\n    float: The sum of a and b.',
    parameters={'properties': {'a': {'title': 'A', 'type': 'number'}, 'b': {'title': 'B', 'type': 'number'}}, 'required': ['a', 'b'], 'type': 'object'}
)

Describe the solution you'd like
It'd be nicer to follow one of the python docstring styles and parse out the argument descriptions from the docstring.

{
  "name": "add_numbers",
  "description": "Adds two numbers and returns the sum.",
  "parameters": {
    "type": "object",
    "properties": {
      "a": {
        "type": "number",
        "description": "The first number to add."
      },
      "b": {
        "type": "number",
        "description": "The second number to add."
      }
    },
    "required": ["a", "b"]
  }
}

Describe alternatives you've considered
i think pydantic-ai does this here:
https://github.com/pydantic/pydantic-ai/blob/10eb5b8523af8e9baae5561f757b2552221c4d0e/pydantic_ai_slim/pydantic_ai/_function_schema.py#L74-L225

[jlowin]

It's an interesting enhancement - I don't know if we'll prioritize at the moment because it involves a lot of new code even if we lean heavily on griffe, unless I'm missing an easy way to export griffe parsing to jsonschema?

[sinamoeini]

seems that openai agent sdk already dos this maybe it can be imported from there https://github.com/openai/openai-agents-python/blob/main/src/agents/function_schema.py#L188

[codingjoe]

Hi there!

I'd be happy to roll up my sleeves for this one. Honestly, I thought this was already implemented. But there's no time like the present.

This could be done without Griffe, but it's an excellent package with a very active community. I, personally, don't see an upside to maintaining a worse parser. Thus, that's the route I'd go when implementing this.

I will get to it next week.

Cheers!
Joe

[codingjoe]

😅 a win is a win, thanks @jlowin