This change doesn't prevent you from moving windows slightly out of
bounds, as long as at least one pixel of the title bar is still on the
screen. That way, you can always grab the title bar and use it to drag
the window somewhere else.
Also, I've made sure that temporarily resizing the window or changing
the GUI scale doesn't ruin your ClickGUI layout. The windows only move
permanently to their new position if you drag them around, otherwise
they move back to their original position once the window is big enough
again.
- Added support for locally hosted models via
oobabooga/text-generation-webui
- Added some more settings (but many are still missing)
- Didn't fix OpenAI mode. It's still hardcoded to the now-deleted
code-davinci-002 model.
This caused more issues than it fixed. If you type, for example: "How
are", the model might respond with " you doing today?". But removing the
whitespace would turn that into "How areyou doing today?".
- Limit the number of suggestions kept in memory and remove duplicates.
- Limit the number of suggestions shown on screen.
- Allow it to generate up to 3 suggestions for the same draft.
This is still unfinished, but it's pretty stable and usable at this
point.
It intentionally has no settings yet, because I might still make changes
at this point that would completely change the way it works. Such as:
- Adding support for OpenAI's ChatGPT models, which need a completely
different prompt
- Adding support for other providers like GooseAI, which have a
completely different API with different parameters, etc.
- Adding support for self-hosted language models like LLaMA-7B, which
have very different parameters again
We might also need some new setting types for this hack, such as a
StringSetting for any custom prompts and an EncryptedStringSetting to
store the API key. The current method of storing the API key in an
environment variable works, but it's not very user-friendly.