Function to Tool Schema

Python module to automatically convert a given Python function into tool schema appropriate for function calling

To execute function calling from GPT responses, we need to define the function and its schema as a GPT tool. However, it can be inconvenient to define both, especially if we need to incorporate multiple functions as GPT tools.

This process can be simplified if we utilize type-hinting and annotations on functions supported by Python using libraries such as inspect and ast. Summary of our schema generation process:

Let us start with a well-documented function get_weather_information:

# Define the function to get weather information
from typing import Optional

def get_weather_information(
    city: str,  # Name of the city
    zip_code: Optional[str] = None,  # Zip code of the city (optional)
):
    """Get weather information for a city or location based on zip code"""
    return {
        "city": city,
        "zip_code": zip_code,
        "temparature": 25,
        "humidity": 80,
    }

---

get_weather_information

 get_weather_information (city:str, zip_code:Optional[str]=None)

Get weather information for a city or location based on zip code

	Type	Default	Details
city	str		Name of the city
zip_code	Optional	None	Zip code of the city (optional)

The function is well-documented and its annotations contain almost all necessary information for tool schema. We will build utilities around extracting such information and converting them into appropriate formats for tool schema.

Parameter descriptions

We can extract the descriptions of function parameters with the ast library. In our implementation, we can follow the inline comments for conveniency. The descriptions can be extracted as follows:

---

source

extract_parameter_comments

 extract_parameter_comments (func:Callable)

Extract comments for function arguments

	Type	Details
func	Callable	Function to extract comments from
Returns	dict	Dictionary with parameter comments

Example of extracting comments from the function:

extract_parameter_comments(get_weather_information)

{'city': 'Name of the city', 'zip_code': 'Zip code of the city (optional)'}

Type converter

Python types cannot be directly transferred into acceptable data types in GPT-compatible tool schema. Therefore, we need an utility to convert these types:

---

source

param_converter

 param_converter (param_type, description)

Convert Python parameter types to acceptable types for tool schema

	Type	Details
param_type		The type of the parameter
description		The description of the parameter
Returns	dict	The converted parameter

Test with parameters of get_weather_information:

city_param = param_converter(str, "Name of the city")
zip_param = param_converter(Optional[str], "Zip code of the city (optional)")
city_param, zip_param

({'type': 'string', 'description': 'Name of the city'},
 {'anyOf': [{'type': 'string',
    'description': 'Zip code of the city (optional)'},
   {'type': 'null',
    'description': 'A default value will be automatically used.'}]})

Function to Schema

We can combine the above utilities with other utilities in inspect to extract information from a Python function and generate a tool schema.

---

source

function_schema

 function_schema (func:Callable, service_name:Optional[str]=None,
                  fixup:Optional[Callable]=None, **kwargs)

Generate a schema from function using its parameters and docstring

	Type	Default	Details
func	Callable		The function to generate the schema for
service_name	Optional	None	The name of the service
fixup	Optional	None	A function to fix up the schema
kwargs	VAR_KEYWORD
Returns	dict		The generated tool schema

Test with our current function:

tool_schema = function_schema(get_weather_information, service_name="Weather Service")
tool_schema

{'type': 'function',
 'function': {'name': 'get_weather_information',
  'description': 'Get weather information for a city or location based on zip code',
  'parameters': {'type': 'object',
   'properties': {'city': {'type': 'string',
     'description': 'Name of the city'},
    'zip_code': {'anyOf': [{'type': 'string',
       'description': 'Zip code of the city (optional)'},
      {'type': 'null',
       'description': 'A default value will be automatically used.'}]}},
   'required': ['city']},
  'metadata': {'module': '__main__', 'service': 'Weather Service'}}}

Simulated GPT workflow

Test integrating with our current GPT framework:

from llmcam.core.fc import *

tools = [function_schema(get_weather_information, service_name="Weather Service")]
messages = form_msgs([
    ("system", "You can get weather information for a given location using the `get_weather_information` function"),
    ("user", "What is the weather in New York?")
])
complete(messages, tools=tools)
print_msgs(messages)

>> System:
You can get weather information for a given location using the `get_weather_information` function
>> User:
What is the weather in New York?
>> Assistant:
The current weather in New York is 25°C with 80% humidity.