Created 05-01-2024, last change 05-01-2024
The OpenAPI descriptions that worked for ChatGPT plugins work immediately with the GPTs. You just have to move the information that was in the ai-plugin.json into the GPT configuration itself. The GPT Builder assistant gives you a good shot at it when you just paste in the ai-plugin.json. Possibly it’s best to configure the actions per hand before doing that.
Returning JSON and text files as response is fine. Returning images so far not .
The files uploaded as knowledge can be used two ways.
If you upload a zip or individual files and enable the code interpreter, the GPT can use that. Likely you can include python programs to execute (didn’t try that yet), but the GPT can create python programs to use the files, though it seems to make sense to give it specific instructions. Probably it makes much sense to include python programs to search for stuff etc. to speed up things. It can show images but as of 1/24 I couldn’t get it to use its visual capabilities on them.
If you don’t enable the code interpreter, the GPT can use the files directly. I tried .md, .txt, .png and .html and .zip files - as of 1/24 it could deal only with .txt and .html files but could easily answer questions from the contained text. I suppose it’s best to include the information you want the GPT to use into one or a few .txt files in markdown format, or perhaps as HTML if that’s what you have (probably that’s more difficult for the GPT).
Remember: a GPT is also just a (well, not quite) human. Give it a good error message, so that it can either recover from the error and retry, or at least tell the user what went wrong. Examples are: