agnaistic
Document the different AI integrations a little bit better
I've been fiddling with the different options, and I feel like there's a lot to learn and a lot of guesswork, understanding how each option compares, and how to configure them, how much it costs, etc. Here's a quick recap of what I figured out up to now. It may be useful to either have a wiki, or include some instructions within the app itself. I'd love to hear from others though, not necessarily about everyone's opinions of the AI itself (we have Reddit for that!) but rather if they have information about setup, pricing, and limitations.
Right now I'm using OpenAI, and I love the idea of switching to another backend in the middle of a conversation.
AI Horde
This is a simple solution but sends your prompts to strangers' machines. You can be fine with it, or not, but it's important to know.
Kobold
Instructions here on how to set it up: https://github.com/KoboldAI/KoboldAI-Client
This can run in Google Colab (simple and easy) or on your machine. All instructions in the link.
For those who never used it, Pygmalion is the go to model for NSFW chats. It's pretty good, but not comparable to giants like GPT-3.5 and GPT-4.
Still, it's a good, safe, and free option.
OpenAI
Login on https://platform.openai.com/ to create an API key, and that's pretty much it. Pricing information here: https://openai.com/pricing. I saw absolutely no filter, which is surprising because their policy prevents usage as an "erotic chat" or "pornography": https://openai.com/policies/usage-policies
Scale Spellbook
My biggest issue is that I was unable to find the pricing for the different models... and the documentation, especially when something goes wrong, is very limited.
Model: GPT-3.5 Turbo and GPT-4
Fairly simple to set up too. Go to https://spellbook.scale.com/ once you have an account, create an app variant with either GPT-3.5 Turbo or GPT-4, type the gaslight prompt in the prompt template, and deploy. In the deploy screen you'll get the URL and key.
Here's what I used, it worked perfectly for me:
System:
Enter roleplay mode. You will write the character's next reply in a dialogue between you and the user. Do not decide what the user says or does. Use Internet roleplay style (no quotation marks, and write your actions in italic in first person like: *I jump in excitement*). You are allowed to use markdown. Be proactive, creative, drive the plot and conversation forward. Write at least one paragraph, up to four. Always stay in character. Always keep the conversation going. Avoid repeating the same ideas from the previous few exchanges.
User:
Description of character:
nice, helpful, cute, easygoing, curious, empathetic, affectionate
Context of the dialogue: The character and the user are talking in a quiet and private area, where they are both comfortable.
This is how the character should talk:
user: Hi!
character: Hi! *I giggle* I'm glad you came! *I tilt my head, smiling warmly* What did you want to talk about?
Continue this conversation:
user: {{ input }}
character:
Models: CoHere, FLAN and AI21
I was unable to make them "behave", maybe with post processing they could work, for example by only reading back the first line, but I wasn't very impressed by how good they were. They do seem to allow NSFW, but only CoHere seems to "play along". Maybe someone who knows better about those models could suggest a way to make them only respond the relevant text and not continue with answering for the user.
NovelAI
This looks good but I didn't try their more advanced options. I also heard that they are working on training a new model, if they pull it off I might try it.
LuminAI
This is a customized version of Kobold AI (Kobold Interceptor) which may instead become a middleware (pre / post processor), TBD.
So, what do I want with this issue?
Documentation, really. But worst case, maybe this issue will be enough for a while :)
What else sounds potentially interesting?
Those would be separate issues, but I think it's interesting to see where this goes... https://goose.ai/ and https://nlpcloud.com/ for example, looks like the models may not be up to par with others but it's cheap. I'll play with those and try to wrap my head around it a little bit more.
There's also a lot of new things to expect from llama and alpaca, such as https://github.com/databrickslabs/dolly though right now it's kind of a mess with this whole tower sitting on llama's "you shouldn't be able to download it" situation...
Note: I just saw that:
- This exists in the repo, which is great since it allows contributions: https://github.com/luminai-companion/agn-ai/tree/dev/instructions
- The home page already has some documentation: https://agnai.chat/ but I'm thinking it might be moved in the Settings page instead?