This repository contains my experiments writing Drafts actions to play with OpenAI's GPT-3 API .
These are the things that work pretty well.
This was my first attempt. It takes the contents of the draft and sends it to the Completion endpoint. The results are appended to the draft. It does some other stuff with selected text, but it's not one of my better ideas.
This is the action that's in the Drafts action directory.
This one uses the Chat endpoint, which requires the prompt to be in a specific chat format. It looks like this:
[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Who won the world series in 2020?"},
{"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
{"role": "user", "content": "Where was it played?"}
]Writing this in JSON is a pain, so in in Drafts you could write it like this.
S: You are a helpful assistant.
U: Who won the world series in 2020?
A: The Los Angeles Dodgers won the World Series in 2020.
U: Where was it played?
OpenAI provides a little guidance on what role values mean,
but they're not very helpful. So, if you leave the role off,
the script assumes user.
All of GPT-3.5's responses are appended to the draft
and prefixed with A:.
S: You are a helpful assistant.
U: Who won the world series in 2020?
A: The Los Angeles Dodgers won the World Series in 2020.
U: Where was it played?
A: The 2020 World Series was played at Globe Life Field in Arlington, Texas.
These are some things I played with.
I was trying understand what the Edit endpoint does. But I either couldn't figure it out or I lost interest.
This was an attempt to generate summaries for websites. Because it doesn't actually visit the sites, it does very well with famous URLs, but makes stuff up when it doesn't.
This was an attempt at an auto-tagger, but then two things happened. First, I had too many tags, and second, it looked hopeless.