Developers can link their Third-party applications with ChatGPT through OpenAI plugins. By interacting with developer-defined APIs, these plugins augment the capabilities of ChatGPT. Chat Plugins enable ChatGPT to do things like:
- Pull up-to-the-minute data, such as live sports results, current stock market values, or the most recent news.
- Access database resources like company documents or personal annotations.
- Provide assistance with tasks such as arranging flight reservations, ordering meals, and more.
Plugins are in a beta and developer access may not be accessible to you. Please join the waitlist to get developer access.
As you delve into the documentation the plugin quickstart repository serves as a useful starting point.
Developers of plugins make available one or more API endpoints, along with a uniform manifest file and an OpenAPI specification. This arrangement enables ChatGPT to digest these files and place calls to APIs as defined by the developer.
The AI model operates as an astute API requester. With an API specification and a natural-language guideline for using the API, the model actively invokes the API to accomplish tasks.
For example, if a user questions:
“Where’s the best place for a couple to stay in Paris?”
The model might opt to reach out to a hotel booking plugin API, receive the response from the API, and craft an answer for the user that melds the API data with its natural language skills.
Plugin flow
Understanding the complete process is crucial when constructing a plugin.
- Start by creating a manifest file and host it on yourdomain.com/.well-known/ai-plugin.json. This file should contain metadata about your plugin (like name, logo, etc.), information about required authentication (type of authentication, OAuth URLs, etc.), and an OpenAPI specification for the endpoints you want to display. The model will encounter the OpenAPI description fields, which can be used to offer a natural language explanation for various fields. Devs advise initially unveiling only 1-2 endpoints with a minimum amount of parameters to reduce the extent of the text. The description of the plugin, API requests, and API responses are all integrated into the conversation with ChatGPT, which contributes to the model’s context limit.
- Register your plugin in the ChatGPT user interface. Choose the plugin model from the top drop-down menu, then click on “Plugins”, “Plugin Store”, and finally “Develop your own plugin”. If authentication is needed, provide an OAuth 2
client_idandclient_secretor an API key. Users need to manually activate your plugin in the ChatGPT UI (your plugin will not be used by default by ChatGPT). You can share your plugin with up to 100 additional users (only other developers can install unverified plugins). If OAuth is needed, users will be redirected via OAuth to your plugin to sign in. - Users initiate a conversation. OpenAI will insert a succinct description of your plugin into a message to ChatGPT, invisible to end users. This will encompass the plugin description, endpoints, and examples. When a user poses a relevant question, the model may choose to initiate an API call from your plugin if it deems it appropriate.
- The model will integrate the results from the API call into its response to the user. It might also include links returned from the API calls in its reply. These will be displayed as rich previews (utilizing the OpenGraph protocol, where we pull the site_name, title, description, image, and url fields).
Markdown
Additionally, the model has the ability to format data from your API in markdown. In this case the ChatGPT UI will automatically render the markdown.