Make module configurable

[TheComamba]

This might be especially interesting if we want to support several languages.

[TheComamba]

Yes, we can. Although it is not large.

[TheComamba]

I think we should offer the user choice of

• The used language model
• The min and max number of tokens
• The prompt framing

[Armandeus66]

If there were an option to send to chat using "/talk" we could integrate this with Talking Actors, which allows us to assign a token in a scene with an AI voice from an Elevenlabs subscription API. https://github.com/acd-jake/acd-talking-actors

[TheComamba]

@Armandeus66 And all we'd have to do is offer the option to add a /talk at the beginning of the reply, correct? That sounds brilliant! Thanks for the input.

[TheComamba]

@Armandeus66 I implemented the functionality with version 0.1.2. Could you verify that it works as expected? I do not have an Elevenlabs subscription. 🙂

[Armandeus66]

Hi! Thanks for being interested in this!

OK, here are my results.

I made an actor and set its Uncanny text to "This character is a typical adventurer." I first made sure that Talking Actors worked properly before creating and using the macro. (You must set it up on a scene, using an actor-linked token, and select the token before typing "/talk Speak this." in chat.)

Both times I executed the macro, asked, "Please introduce yourself." and got a response. The first chat AI could introduce itself briefly, and the second told me an irrelevant story and was cut off after a character limit seemed to have been reached. On the first try, before getting a response, my browser froze for about 15 seconds when using the first AI, and about 5 using the second.

However, the "/talk" was output as part of the chat message (meaning you could see it in the text), not as a chat command, and so the integration with Talking Actors was unsuccessful. (output = /talk I am also pretty good at writing fiction and short stories, too.) I copied the text and saw that sometimes there was a lot of empty space after /talk.

Also, there were what appeared to be some character code problems with the first AI. The first characters of a word were switched to Chinese/Japanese characters, and the period was 2-byte Japanese also. (output = 我 nwere some funny memories of my childhood that were very much alive today。) Another thing is that it appended text emoji such as :)) which would affect what the Elevenlabs might say. (output = If you'd ike see it again soon you'll be glad you're here anyway! :)) )

The cheapest Elevenlabs subscription that gives you an API key is $5 a month, so it's not expensive just to test it for one month and quit.

[TheComamba]

Thanks for the report!

Yeah, the first model is a bit of a weirdo. I can confirm that it takes somewhat longer to generate Tokens, and they often contain non-Ascii characters and bizarre content. Good thing the model is configurable now.^^

I have the growing suspicion that the way I generate the chat messages somehow bypasses the text macro handling of foundry. I will investigate.

But the model is configurable now, so I will track this in the new issue #27.