English | 中文
A minimally framework for creating a function that can be invoked by OpenAI's function calling mechanism, written in NodeJS and deployed on Vercel, tested by ConsoleX.ai for function invocation.
We will use an example of creating a function to answer questions about the 2022 FIFA World Cup (since the training data cutoff for ChatGPT models is currently January 2022, calling a function is needed to give correct answers). With simple steps, you can complete the creation, deployment and testing of the function.
First, clone the code locally:
git clone https://github.com/quentinzhang/OpenAI-function-boilerplate.git
Next, enter your project directory
cd OpenAI-function-boilerplate
If you have not installed Vercel, first install it globally:
npm i -g vercel
Then initialize a Vercel project in the root directory of the project:
vercel
Finally, deploy the project to production:
vercel --prod
After the operation is complete, your production domain name should be displayed in the terminal:
https://<your-vercel-domain>
At this point, you have completed the serverless function deployment process. Your function invocation path is:
https://<your-vercel-domain>/api/fifa_worldcup_2022_info
Since this function does not require any parameters, you can see the JSON information returned by the function by entering it in a browser.
The description of the function is contained in the function_description.json file, which can be passed as a parameter defining the function when invoking OpenAI's API.
Debugging function invocation can be cumbersome as it involves several intermediate steps. Below is a shortcut to test function invocation with ConsoleX.ai.
Open the function_description.json file, copy the description information about the fifa_worldcup_2022_info function, and paste it into ConsoleX.ai's function description configuration and click the Save button.
Then, copy your production domain name, append /api to it as: https://<your-vercel-domain>/api. Fill it in as the base path configuration on ConsoleX.ai's homepage.
Next, turn on the switch to enable OpenAI function calling, and ask GPT a simple question about the 2022 World Cup:
Which team won the 2022 FIFA World Cup?
You should now be able to see the correct answer returned by OpenAI's API through function invocation.
You can develop new functions in a similar way - simply replace the function description information in function_description.json with your own, and write the function implementation code under the api directory in the corresponding file.
You can also refer to the Vercel documentation for more information on how to develop serverless functions.
The example also contains implementation code for a get_current_weather function. You can refer to it to implement a function that needs to call third-party RESTful API interfaces.
Install the axios dependency first by running npm install, and add an environment variable named OPEN_WEATHER_MAP_API_KEY in Vercel's project settings, and fill in the API Key you applied from openweathermap as the value.
During local development, you can copy the .env.example file to .env, and fill in your API Keys in it. This allows you to use it when debugging locally.
cp .env.example .env
If you want to encrypt the function you are developing, you can set the value of the environment variable EXPECTED_API_KEY to a random string, and then pass that string in the authorization bearer token in the header when calling the function. This ensures that only those who know that string can invoke your function.