BenzGPT is am AI Agent that uses the PromptAPI to understand a given command and then decide what it should do next. Should it just chat? Or should it move?
More than that it also uses the WebSpeechAPI to transcribe spoken words into text and then uses the SpeechSynthesisAPI to speak back to the user. So you can actually talk to BenzGPT!
To round it all off, of course, it's not just a virtual conversation. A toy car can also be controlled via WebBluetooth.
The core of BenzGPT is the FunctionCallingPromptAPI. It is an abstraction layer to convert any LLM (in this case the PromptAPI) into a function calling API.
const benzGPT = new FunctionCallingPromptAPI();
// register a function
benzGPT.registerFunction<number>({
name: 'move',
description: 'Sets the new speed of the car',
parameter: {
name: 'speed',
description:
'can be a number between -100 (full speed backwards), 0 (stand still) and 100 (full speed)',
type: 'number',
},
returnType: 'void',
examples: [
{
message: 'I am moving forward with full speed, speed is set to 100%',
function: 'move',
parameter: 100,
},
{
message: 'I did low down a bit. My new speed is 50%',
function: 'move',
parameter: 50,
},
{
message:
'I am moving backward with full speed, speed is set to -100%',
function: 'move',
parameter: -100,
},
],
run: runMove,
});
// initialize the session
await benzGPT.initializeSession(
callback,
'You are an intelligent car named "BenzGPT"'
);
// generate a response
const response = await benzGPT.generateResponse('I would like you to go full speed');The response will then always be an object with the following structure:
{
content: string;
role: 'system' | 'user' | 'assistant';
error?: string;
parsed: {
message: string;
function: string;
parameter: string | number;
functionResult?: any;
}
}If a function should be called, the function and parameter will be set. If the function was called successfully, the functionResult will be set.
registerFunction is a method that accepts a function definition in a structured way and then generates a system prompt for the LLM that can then allow the LLM to use this function.