Talk with GPT 3.5 model right in Discord. Powered by OpenAI's Whisper & GPT.
Well, your voice is sent to the local Whisper backend to get text which adds to the room context.
Aikashka will respond every 10-40 seconds or by calling her(/him) by name.
When the time comes, a big prompt is being sent to the OpenAI's GPT endpoint. The answer will be strictly formatted. Depending on the response, Aikashka can:
- say something
- write something in the text channel
- kick someone from the voice channel
- save some note for herself
- leave the channel
Note that she remembers only 20 last user messages and 6 her own responses, all in order.
I don't do pre-made binary. It means that you have to clone this repo, open it in JetBrains Rider / Fleet (or Visual Studio (/Code) if you don't respect yourself) and make some changes on your own.
TTS made using Tinkoff VoiceKit (well, it's really cheap and the voice is REALLY good).
Rename appsettings.example.json to appsettings.json and fill all required API keys.
Rename name_mappings.example.txt to name_mappings.txt and map all user profiles that will talk with the bot.
Take a look at systemMsg in Aikashka.Core/Generators/TextGenerator.cs - first 3 lines. Change them if you want.
Create virtual environment & install requirements.txt.
If you want GPU acceleration, install PyTorch with CUDA, e.g. pip install torch --index-url https://download.pytorch.org/whl/cu117
And if you don't, then remove , device=0 from core.py
Also ensure that model fits your needs.
Compile & run bot.
Invite bot to the server and use /go command to join the voice channel.
/go: forces bot to join the channel./cum: send message to the room context. useful if you can't spell properly some words/state: shows room context./clear: clear room context.