function_calling.md

Function Calling

Define a function:

Define a function (a standard C# method), I create an "Add" method here which adds two numbers and returns the result, optionally changing the base:

(Note that the method returns string, this is because the result of a function call needs to be read by the model (as a string). This applies to all functions provided to the model. The model will intepret the return value however it wishes, whether that be int, floating-point, string, etc)

private static string Add(int a, int b, int toBase = 10)
{
    // Code goes here...
}

Decorate the method:

Decorate the method with the relevant attributes FunctionDescription, ParamDescription, and (if relevant) EnumValues (more on this); which will enable the ChatGPT class to parse the method.

Further reading (expand)

The ChatGPT class parses the method into a FunctionDefinition prior to sending it to the model. A FunctionDefinition is a serializable object which the model recognises)

[FunctionDescription("Adds 'b' to 'a'.")]
private static string Add([ParamDescription("The first value", true)] int a,
                          [ParamDescription("The second value", true)] int b,
                          [ParamDescription("What base the output should be in", false), EnumValues(2, 10, 16)] int toBase = 10)
{
    // Code goes here...
}

Alternative to decorating a method (expand)

You can create the FunctionDefinition manually:

FunctionDefinition additionFunction = new FunctionDefinition.Builder("Add", "Adds 'b' to 'a'.")
  .AddParameter("a", "integer", "The first value", true)
  .AddParameter("b", "integer", "The second value", true)
  .AddParameter("toBase", "What base the output should be in", false, new int[] { 2, 10, 16 })
  .SetMethod(Add)
  .Build();

This is functionally identical to decorating the method.

Call the model:

Call the ChatCompletions API with your function:

(Note: you must use the 0613 variants of both GPT 3.5 Turbo and GPT-4 as of 7th July 2023 because the default models do not currently support functions)

var response = await gpt.GetChatCompletion(
                          messages: new List<Zintom.OpenAIWrapper.Models.Message> {
                            new() { Role = "user", Content = "What is 9 + 900?" }
                          },
                          options: new ChatGPT.ChatCompletionOptions() { Model = LanguageModels.GPT_3_5_Turbo_0613 },
                          functions: Add);