Enable Azure Open AI function calling via Azure Functions

[goventur]

Motivation and Context

Description

This PR implements OpenAI Function Calling feature.
It uses Azure Functions to retrieve the OpenAI tools metadata JSON, and also to run the functions.
To add new OpenAI Functions, only the code in Azure Functions needs to be modified.

Architecture

1, 2 - Get JSON with OpenAI function call metadata
3, 4, 5, 6 - Send user prompt to Open AI model optionally with On Your Data, querying Azure AI Search
7, 8 - If Open AI model returns any function call, execute the function remotely using App Functions
9, 10 - Submit the results from Azure Function call execution to OpenAI to generate the final answer

This enables users to add real-time data to their chatbots, or to implement actions.

Azure Functions

There are two Azure Functions:

1. /tools: returns JSON metadata with the available OpenAI functions.
2. /tool: executes the OpenAI function with the given arguments.

Azure Functions code sample

import azure.functions as func
import logging
import json
import random

app = func.FunctionApp(http_auth_level=func.AuthLevel.FUNCTION)

azure_openai_tools_json = """[{
    "type": "function",
    "function": {
        "name": "get_current_weather",
        "description": "Get the current weather in a given location",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {
                    "type": "string",
                    "description": "The city name, e.g. San Francisco"
                }
            },
            "required": ["location"]
        }
    }
}]"""

azure_openai_available_tools = ["get_current_weather"]

@app.route(route="tools")
def tools(req: func.HttpRequest) -> func.HttpResponse:
    logging.info('tools function processed a request.')

    return func.HttpResponse(
        azure_openai_tools_json,
        status_code=200
    )

@app.route(route="tool")
def tool(req: func.HttpRequest) -> func.HttpResponse:
    logging.info('tool function processed a request.')

    tool_name = req.params.get('tool_name')
    if not tool_name:
        try:
            req_body = req.get_json()
        except ValueError:
            pass
        else:
            tool_name = req_body.get('tool_name')

    tool_arguments = req.params.get('tool_arguments')
    if not tool_arguments:
        try:
            req_body = req.get_json()
        except ValueError:
            pass
        else:
            tool_arguments = req_body.get('tool_arguments')

    if tool_name and tool_arguments:
        if tool_name in azure_openai_available_tools:
            logging.info('tool function: tool_name and tool_arguments are valid.')
            result = globals()[tool_name](**tool_arguments)
            return func.HttpResponse(
                result,
                status_code = 200
            )

    logging.info('tool function: tool_name or tool_arguments are invalid.')
    return func.HttpResponse(
            "The tool function we executed successfully but the tool name or arguments were invalid. ",
            status_code=400
    )

def get_current_weather(location: str) -> str:
    logging.info('get_current_weather function processed a request.')
    temperature = random.randint(10, 30)
    weather = random.choice(["sunny", "cloudy", "rainy", "windy"])
    return f"The current weather in {location} is {random.randint(10, 30)}°C and {weather}."

Contribution Checklist

• I have built and tested the code locally and in a deployed app
• For frontend changes, I have pulled the latest code from main, built the frontend, and committed all static files.
• This is a change for all users of this app. No code or asset is specific to my use case or my organization.
• I didn't break any existing functionality 😄

[sarah-widder]

can we make this Optional[bool] for consistency?

[sarah-widder]

let's use an async call here

[sarah-widder]

let's make this async as well

[sarah-widder]

a safety check here that len(messages) > 0 would be good