Implementing Local LLMs in TeaseAI

[Mr_kinky]

Hello everyone,

I've been playing around with local LLMs, i.e. language models, a lot lately. There are now some that even run smoothly on my laptop and are not censored, so you can have really kinky conversations with the AI (Like Mag Mell on Huggingface).
Some time ago I saw that a script was written for sexscript that accesses the local server from the local LLM to generate the answers in this way, i.e. not to give any ready-made answers at all, but only hints and then have the generated answer returned to the program.

https://ss.deviatenow.com/viewtopic.php?f=7&t=1116

I find the topic incredibly exciting, but fail because I can't program properly, I can't even adapt existing code to implement something like this. Therefore, this is more of a brainstorm post that can be seen as a suggestion.

I have LM Studio installed on my laptop, which has an integrated server.

Now I had the idea of creating a version of TeaseAI that not only returns the texts from the files, but also sends a request to the local server each time with information from the files, possibly also already defined parameters, such as session duration, preferences or anything else, like the flags in the original program.

The advantage would be that everyone can customize their own personality and scenarion, because the important files are available in TeaseAI as text files that can be easily edited.

So, that's the input for now. Unfortunately, I don't know how time-consuming a query to the local LLM server is, nor how time-consuming an implementation in the software or the rewriting of suitable software is. Maybe this approach is not interesting for others, but I wanted to get rid of it.

Greetings

[markus]

The advantage would be that everyone can customize their own personality and scenarion, because the important files are available in TeaseAI as text files that can be easily edited.

Hi!

Mind to elaborate a bit more on this point?
Did you played with personalities from TAI?

What I mean, ... the customize part, ... what I have in mind is more of what you have described before, like the flags (i.e. what fetishes), values (how long denied already), maybe a describtion of the character of the mistress (gentel, playful, strict, etc).
I am just brainstorming, too!
So when I think about that, there wouldn't be a customize part, because the request towards the server would be constructed by what is given by the personality and the values.

So, what you have in mind to customize?

I find the topic incredibly exciting

You are not the only one.


I've wrote a little already about this here (see also some posts above for where this started):
viewtopic.php?p=368872#p368872

Best greetings,
Markus

[Daragorn]

Having an AI driven tease AI has always been my dream program too.
I have been trying to re-create a similar experience with all AIs out there to no avail, there is no way to get any of them "lead" a session credibily and coherently; they keep forgetting things and/or lose tracks and derail, so it ends up you pratically have to "suggest" them the orders. On top of that, there are no slideshow going on like on TAI that keeps you hooked to the session and/or sometimes are tied to a specific script/scenario, no beatmeter, etc etc.

TLdR: if someone manages to find a way to have TAI generate resposnes through an AI, but also remembering the flags/status/duration/etc etc (and, even better, controlling also the metronome and such), it would be absolutely awesome.

[avatarbr]

Well, my idea on Markus link was something like that.

But the LLM would not 'lead the tease', would just add more variety. Something to be active when the user type something, and the Personality don't have a response for that.

The harder part, without considering the implementation, would be making the LLM 'know' what is happening. Yes, you can setup for a teasing session, but the LLM need to know if you are stroking, edging, how long is the session, how many edges you already did or anything else.

Well, a lot of these things are stored in variables, but I don't know how you could use it to tell the LLM what is going on.

Another thing is, not always the LLM will make sense, so we still need the Personality to be in charge. Imagine if the LLM tell you to edge 200x. So the personality will need to have the parameters to control these things.

[Mr_kinky]

Hi and thanks for the replies,


@markus I didn't see your post, it sounds similar to my wish :D

And just like Daragorn, I also have the problem with pure chat sessions that they don't really run smoothly from start to finish.

On the simple editing of personalities: I've played through a few in Tease AI and found it easy to customize them to my liking in that I could change the text and commands in the txt files, or just add and retrieve flags, for example. I envision such a modular structure and easy customizability for the LLM controlled version as well.

In principle, I was thinking of the following, but it is not yet fully developed:
In the settings, you can set the setting in the form of prompts and define the mood and other factors. These are then processed internally in a prompt.
The basic structure should remain the same as with Tease AI, i.e. start files, session files, end files, maybe even different categories for different fetishes.
Also the time in the background and the sequence, so that the session simply remains within a framework.

But the txt files are retrieved, a prompt is formed from them and the llm then formulates a text. For example, duration of the session, preferences, what was decided, what was discussed, what should happen in this module. And then the module remains in the context until the txt is finished and tease ai wants to edit the next txt again. When a file is "finished", flags can be updated here, a summary of the module can be written, and so on. Then the context is empty again and the context is regenerated for the next txt file with the summaries of the session, the preferences and the information about what should happen in this module.

So basically clean up the old tease ai (or make a fresh start) and look how te interact with a local server (like from the Sexscript script from above) and then try how it works when you just use the llm to generating the chatting text and handle the organisation inside tease ai.

But again, i love to try and look into the functionality of the llms, the prompts, the txt files, but dont know anything about the actual coding that is needed to make an actual program..

[Mr_kinky]

ok, well, i only see avatarbr's message now :D

So I was thinking, as perhaps mentioned above, of filling a text file during the session that summarizes the session so that the llm always knows what has already happened based on this file.
At the end, the file can be used to show the llm what happened in the last (and penultimate and so on) session