Function calling empowers users to extend language model interactions, seamlessly integrating custom code for dynamic, application-specific logic. Recent models like gpt-3.5-turbo-1106 and gpt-4-1106-preview are adept at identifying when a function should be invoked.
GPTBoost logs offer convenient statistics tracking for prompts finishing in a function call. Users can effortlessly filter requests invoking specific functions, enhancing visibility and control over their tailored conversational experiences.
Evaluate function usage with GPTBoost
Discover how your functions are utilized in real time through GPTBoost's intuitive dashboard. Each interaction with the language model is logged, allowing you to effortlessly track and analyze the frequency and performance of your functions. Dive into essential metrics like calls, tokens, and costs associated with each function.
Effortlessly examine the distribution of function calls with just a click on the 'Browse Requests' button. GPTBoost puts the power of detailed analysis and informed decision-making at your fingertips, providing valuable insights into the real-world usage of each function.
Recommended Steps in Function Calling
In the API request, you can define functions, and the model can intelligently decide to generate a JSON object containing arguments for calling one or multiple functions. It's important to note that the Chat Completions API itself doesn't execute the function; instead, it produces JSON output that you can utilize to invoke the function within your code. OpenAI recommends the following workflow.
Invoke the Model with Functions: Call the model by providing the user query and a set of functions defined in the functions parameter.
Check if a function call is needed - The model might choose to execute one or more functions, resulting in content structured as a JSON object following your custom schema (bear in mind that the model may create parameters).
Parse JSON and Execute Functions: Parse the JSON string obtained from the model's response and execute your custom functions, unpacking the provided arguments.
Trigger the Model Again: Invoke the model once more by adding the function response as a new message, enabling the model to provide a summary of the outcomes to the user.
Code Examples
Have a look at the following code examples showcasing how to leverage function calling. Make sure to incorporate necessary checks and exception handling tailored to your specific use case to ensure robust and reliable functionality.
# This example is for v1+ of the openai: https://pypi.org/project/openai/
import os
import json
import requests
from openai import OpenAI
client = OpenAI(
# GPTBoost API base URL
base_url= "https://turbo.gptboost.io/v1",
api_key= os.getenv('OPENAI_API_KEY'),
)
# Weather API credentials
weather_api_url = "https://weatherapi-com.p.rapidapi.com/current.json"
headers = {
"X-RapidAPI-Key": os.getenv('RAPIDAPI_KEY'),
"X-RapidAPI-Host": "weatherapi-com.p.rapidapi.com"
}
# Define function to check the weather for a specific location
def get_current_weather(city: str):
querystring = {"q":city}
response = requests.get(weather_api_url, headers=headers, params=querystring)
# Return the response as str to pass in the next request to the LLM
return response.text
# The prompt from the user
messages = [{"role": "user", "content": "What's the weather in Maldives"}]
# The function info that will be passed to the LLM
functions = [
{
"name": "get_current_weather",
"description": "Get the current weather in a given city",
"parameters": {
"type": "object",
"properties": {
"city": {
"type": "string",
"description": "The name of the city"
},
},
"required": ["city"],
}
}
]
# Make the initial request to the LLM
response = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=messages,
functions=functions,
function_call="auto"
)
# Check if a function call is needed
if response.choices[0].message.function_call:
function_name = response.choices[0].message.function_call.name
arguments = response.choices[0].message.function_call.arguments
available_functions = {
"get_current_weather": get_current_weather,
}
function_to_call = available_functions[function_name]
# Invole the function with arguments
function_response = function_to_call(**json.loads(arguments))
# adding assistant response to messages
messages.append({
"role": "function",
"name": function_name,
"content": function_response,
})
response_after_function = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=messages,
functions=functions,
function_call="auto"
)
# Print the summarized response to the user
print(response_after_function.choices[0].message.content)
else:
print(response.choices[0].message.content)
// This code is for v4+ of the openai package: npmjs.com/package/openai
import OpenAI from 'openai';
import axios from 'axios';
const openai = new OpenAI({
apiKey: process.env.OPENAI_API_KEY,
baseURL: "https://turbo.gptboost.io/v1",
})
// Define function to check the weather for specific location
async function get_current_weather(location) {
const options = {
method: 'GET',
url: 'https://weatherapi-com.p.rapidapi.com/forecast.json',
params: {
q: location,
days: '3'
},
headers: {
'X-RapidAPI-Key': process.env.RAPIDAPI_KEY,
' X-RapidAPI-Host': 'weatherapi-com.p.rapidapi.com'
}
};
try {
const response = await axios.request(options);
let weather = response.data;
const weatherForecast = `Location: ${weather.location.name} \
Current Temperature: ${weather.current.temp_f} \
Condition: ${weather.current.condition.text}. \
Low Today: ${weather.forecast.forecastday[0].day.mintemp_f} \
High Today: ${weather.forecast.forecastday[0].day.maxtemp_f}`;
return weatherForecast;
} catch (error) {
console.error(error);
return "No forecast found";
}
}
// The user input
const messages = [{ role: "user", content: "What's the weather in Thessaloniki" }]
// The function that checks the weather that would be added to the request to the LLM
const functions = [{
name: "get_current_weather",
description: "Get the weather forecast for a given location",
parameters: {
type: "object", // specify that the parameter is an object
properties: {
location: {
type: "string", // specify the parameter type as a string
description: "The location, only city name e.g. Beijing, China. location: Beijing"
}
},
required: ["location"] // specify that the location parameter is required
}
}
]
// Define a function to interact with the LLM
async function ask_gpt(){
const completion = await openai.chat.completions.create({
model: "gpt-3.5-turbo-16k",
messages: messages,
functions: functions,
function_call: "auto"
})
// Check if a function call is needed
if (completion.choices[0].message.function_call){
// Get all the arguments ready.
const function_name = completion.choices[0].message.function_call.name;
const function_arguments = JSON.parse(completion.choices[0].message.function_call.arguments);
const available_functions = {
"get_current_weather": get_current_weather,
}
const function_to_call = available_functions[function_name]
// Call the function
const current_weather = await function_to_call(function_arguments.location);
// Add the function response to the messages for the next prompt to the LLM
messages.push({
role: "function",
name: function_name,
content: current_weather,
})
try {
const completion_after_function = await openai.chat.completions.create({
model: "gpt-3.5-turbo-0613",
messages: messages
});
console.log(completion_after_function.choices[0].message.content);
} catch (error) {
if (error.completion_after_function) {
console.log(error.completion_after_function.status);
console.log(error.completion_after_function.data);
} else {
console.log(error.message);
}n
}
}
else {
const completion_text = completion.choices[0].message.content;
console.log(completion_text);
}
}
ask_gpt()
# This example is for v1+ of the openai: https://pypi.org/project/openai/
import os
import json
import requests
from openai import AzureOpenAI
client = AzureOpenAI(
# Set GPTBoost API base URL for azure_endpoint
azure_endpoint = "https://turbo.gptboost.io/v1",
api_key = os.getenv('AZURE_API_KEY'),
api_version = "2023-07-01-preview",
)
deployment_name="gpt-35-turbo-16k"
weather_api_url = "https://weatherapi-com.p.rapidapi.com/current.json"
headers = {
"X-RapidAPI-Key": os.getenv('RAPIDAPI_KEY'),
"X-RapidAPI-Host": "weatherapi-com.p.rapidapi.com"
}
# The user input
messages = {"role": "user", "content": "Tell me the wind in Lemnos"},
functions=[
{
"name": "get_current_weather",
"description": "Get the current weather for a given city",
"parameters": {
"type": "object",
"properties": {
"city": {
"type": "string",
"description": "Pass the city to obtain the current weather"
},
},
"required": ["city"]
}
}
]
def get_current_weather(city: str):
querystring = {"q":city}
response = requests.get(weather_api_url, headers=headers, params=querystring)
# Return the response as str to pass in the next request to the LLM
return response.text
response = client.chat.completions.create(
model=deployment_name,
messages=messages,
functions=functions,
function_call="auto",
)
response_message = response.choices[0].message
# Check if the model wants to call a function
if response_message.function_call:
# Call the function. The JSON response may not always be valid so make sure to handle errors
function_name = response_message.function_call.name
available_functions = {
"get_current_weather": get_current_weather,
}
function_to_call = available_functions[function_name]
function_args = response_message.function_call.arguments
function_response = function_to_call(**json.loads(function_args))
messages = [x for x in messages]
# Add the function response to the messages
messages.append(
{
"role": "function",
"name": function_name,
"content": function_response,
}
)
# Call the API again to get the final response from the model
second_response = client.chat.completions.create(
messages=messages,
model=deployment_name
# optionally, you could provide functions in the second call as well
)
print(second_response.choices[0].message.content)
else:
print(response.choices[0].message.content)
Function Usage in GPTBoost Dashboard
Press 'Browse Requests' to check all requests where a specific function was called
