Vulcan

Making a Signal Messenger-enabled chatbot, powered by LLMs on commodity hardware

Structure:

flowchart TD
    Middleware --> C[LLM Handler]
    Middleware --> D[Signal Messenger API]
    Middleware --> E[Emotional State Machine]
    Middleware --> F[NLP]
    Middleware --> G[sampleMessages]
    C --> J[Llama 13b 8-bit quantized model]

At the core of this project are dockerized deployments of the Signal Messenger API, and Llama.cpp server. The middleware is custom code to glue it all together.

Chatbot Personality:

The "personality" is defined by prompt injection in the LLM handler file. Currently Vulcan is set to act as a robotic pirate.

Emotional State:

The emotional state is managed by a state machine (defined below).

• If he gets a rude message, it'll have a chance to get 'Angry' or 'Sad'.
• If he recieves a compliment, his mood improves.
• If he gets too many boring messages in a row, he gets "bored" and asks a question.

Can I use this for my own projects?:

Absolutely! You'll probably need to make changes to the configuration though

Sample Messages:

Next steps:

• State machine to simulate "moods"
• Personality tuning
  • Allow high reputation conversational partners to influence the personality
• Proper logging of messages
• Fine-grained control of which Signal chats to respond to
  • Reply to new chats by default, on wakeword
• Memory
  • Short term/conversational memory
  • Long term/database memory