callforge-ai 0.1.0

Reliable schema-based function calling for local language models.

CallForge

Reliable schema-based function calling for local language models.

CallForge selects a declared function, extracts type-safe parameters from a natural-language request, and uses constrained decoding only when a choice is ambiguous. It is designed for local models and works with any model that implements two small methods:

encode(text)
get_logits_from_input_ids(input_ids)

Installation

Install the lightweight core:

pip install callforge-ai

Install the ready-to-use Hugging Face adapter:

pip install "callforge-ai[huggingface]"

Python usage

from callforge import FunctionCaller, Function
from callforge.adapters.huggingface import HuggingFaceModel

functions = [
    Function.model_validate(
        {
            "name": "fn_add_numbers",
            "description": "Add two numbers together.",
            "parameters": {
                "a": {"type": "number"},
                "b": {"type": "number"},
            },
            "returns": {"type": "number"},
        }
    )
]

model = HuggingFaceModel("Qwen/Qwen3-0.6B", device="cpu")
caller = FunctionCaller(model=model, functions=functions)
result = caller.call("What is the sum of 2 and 3?")

print(result.model_dump())
# {'prompt': 'What is the sum of 2 and 3?',
#  'name': 'fn_add_numbers',
#  'parameters': {'a': 2.0, 'b': 3.0}}

JSON directory usage

Input directory:

data/input/
├── function_calling_tests.json
└── functions_definition.json

Run:

callforge \
  --input data/input \
  --output data/output \
  --model Qwen/Qwen3-0.6B \
  --device cpu

The result is written to:

data/output/function_calling_results.json

Use another local model

You do not need Transformers. Pass any object with the required interface:

class MyModel:
    def encode(self, text: str):
        ...

    def get_logits_from_input_ids(self, input_ids: list[int]) -> list[float]:
        ...

caller = FunctionCaller(model=MyModel(), functions=functions)

Development

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
pytest
python -m build
python -m twine check dist/*

Security note

Only enable --trust-remote-code for model repositories you trust.

License

MIT