2024 Apr 5
ChatGPT Plugins enable ChatGPT to access real-time information. Web service API reference documentation plays a big and important role in the ChatGPT Plugin architecture. You can’t force ChatGPT to use your plugin. You can only persuade ChatGPT to do so by documenting your API effectively.
This post is part of my ongoing series to explore the potential impact of generative AI on technical writing.
(ChatGPT Plugins are in “Limited Alpha” which means that OpenAI might still make fundamental changes to the ChatGPT Plugin architecture. In other words, this post might be completely outdated in a year.)
ChatGPT is an AI chatbot that can generate detailed answers across many domains. The large language model that powers ChatGPT was trained on data from 2021 so ChatGPT can’t provide information about stuff that has happened recently. ChatGPT Plugins fill in this information gap by enabling ChatGPT to access web service APIs. In other words, plugins enable ChatGPT to get post-2021 information.
Here’s the fascinating bit for technical writers. From Introduction:
The AI model acts as an intelligent API caller. Given an API spec and a natural-language description of when to use the API, the model proactively calls the API to perform actions. For instance, if a user asks, “Where should I stay in Paris for a couple nights?”, the model may choose to call a hotel reservation plugin API, receive the API response, and generate a user-facing answer combining the API data and its natural language capabilities.
In other words, you can’t control when ChatGPT uses your plugin. You can only maximize the chance that ChatGPT uses your plugin by describing your API effectively!
Plugin flow gives you a peek into how ChatGPT plugins are built. You create a
plugin manifest at <domain>/.well-known/ai-plugin.json where <domain> is
just a placeholder for your actual domain. The plugin manifest contains a
description of the overall web service API as well as descriptions for each API
endpoint. The endpoint descriptions must conform to the OpenAPI specification.
There are two scenarios to watch out for:
ChatGPT stays popular and becomes a fundamental tool
Some other AI-powered tool becomes fundamental and adopts the ChatGPT Plugins architecture
Effective API reference documentation will probably become much more closely tied to organization success and much easier to measure. A high-quality API reference that closely matches the vocabulary of users and is easy for ChatGPT to consume should result in increased plugin usage.
Technical writers who have lots of experience with the OpenAPI specification in particular should see continued strong demand for their very particular set of skills.
As hinted at in the last section, when writing API reference documentation for ChatGPT, your main audiences are ChatGPT users and ChatGPT itself. This is a big change in focus for technical writers. Currently, we optimize API reference documentation for human developers.
Let’s start with ChatGPT users. Remember that the name of the game is to persuade ChatGPT to call your plugin. Presumably, the more your API reference matches the vocabulary of your customers (the people using ChatGPT), the higher the chances that ChatGPT will recognize that your plugin is a match for the task at hand.
(Emphasis on “persuade” in the last paragraph because OpenAI makes it very clear in Best practices that the API reference must be accurate and objective and must not attempt any manipulation. Plugin flow says that plugins will only be accessible to all ChatGPT users after passing a review.)
Next, ChatGPT itself. I don’t have much to say here, other than the fact that the plugin manifest must be structured and written in a way that makes it easy for ChatGPT to consume your API.
Finally, human developers. As far as plugin manifests are concerned, I don’t really see much of a need to optimize for human developers. Maybe we learn eventually that an API reference that is easy for a human developer to read also happens to be what’s easiest for ChatGPT to consume.
Along with “prompt engineer” I wouldn’t be surprised if “LLMO” (large language model optimization) becomes a trendy new tagline on LinkedIn.
The API reference section of the plugin manifest must conform to the OpenAPI
specification. The .well-known part of the example plugin manifest URL that
you saw earlier is an IETF standard: Defining Well-Known URIs