[Bug]: Tool schema marks **kwargs as a required (untyped) parameter, forcing the LLM to fill it

[gautamvarmadatla]

Bug Description

FunctionTool builds its parameter schema with create_schema_from_function, which loops over a function's parameters without checking param.kind, so *args and **kwargs get added as required schema properties (and unannotated **kwargs has no type). This makes the LLM-facing tool spec demand a phantom args/kwargs argument on every call and gets rejected by OpenAI strict function calling.

Version

0.14.23

Steps to Reproduce

from llama_index.core.tools import FunctionTool
from llama_index.llms.openai import OpenAI


def search(query: str, **kwargs) -> str:
    """Search the web."""
    return "results"


tool = FunctionTool.from_defaults(fn=search)
llm = OpenAI(model="gpt-5.5", strict=True)  

resp = llm.chat_with_tools([tool], user_msg="search for cats")
print(llm.get_tool_calls_from_response(resp, error_on_no_tool_call=False))

Relevant Logs/Tracebacks

ERROR: BadRequestError
Error code: 400 - {'error': {'message': "Invalid schema for function 'search': In context=('properties', 'kwargs'), schema must have a 'type' key.", 'type': 'invalid_request_error', 'param': 'tools[0].function.parameters', 'code': 'invalid_function_parameters'}}

[dosubot]

This is a valid bug. Looking at create_schema_from_function in tools/utils.py, the loop iterates over all parameters from signature(func).parameters without checking param.kind, so VAR_KEYWORD (**kwargs) and VAR_POSITIONAL (*args) are treated like regular parameters. Since **kwargs has Parameter.empty as its default, it gets marked as a required field with type Any — which then fails OpenAI's strict mode validation because Any has no JSON Schema type key.

Workaround: You can use FunctionTool.from_defaults with the tool_metadata parameter to supply a custom fn_schema that excludes kwargs, or wrap your function without **kwargs:

from llama_index.core.tools import FunctionTool

def search(query: str) -> str:
    """Search the web."""
    return _search_impl(query)

tool = FunctionTool.from_defaults(fn=search)

Fix: The proper fix in create_schema_from_function would be to skip variadic parameters:

import inspect

for param_name in params:
    if param_name in ignore_fields:
        continue
    if params[param_name].kind in (
        inspect.Parameter.VAR_POSITIONAL,
        inspect.Parameter.VAR_KEYWORD,
    ):
        continue
    # ... rest of the loop

This is a straightforward oversight — the function just needs a param.kind check to skip *args/**kwargs before building the schema fields.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.