Khatchig-Bot 1.0 was a function-calling, UI-changing chatbot that leveraged Instructor and Pydantic on the back-end.
The json generated by the Large Language Model wasn't consistently parsable due to inconsistent quotes " '.
OpenAI's "json mode" (with Instructor) may be somtething to explore in the future.
Khatchig-Bot 1.1 was a Neural Cellular Automata chatbot.
Inference time at each step was a little slow
on our $7/month 0.5 CPU hobby server, so we temporarily retired him.
Khatchig-Bot 1.2 was a Particle Affinity chatbot run over a websocket w/ dependency injections for user state management on the back-end.
The next state computation was faster than the NCA inference step, but it was slow serving multiple people on the 0.5 CPU server.
Khatchig-Bot 1.3 is also a Particle Affinity chatbot, but the particle simulation is now run on the client side.
As a result, the back-end is much simpler, and the simulation details are effectively open-sourced in JavaScript.
đž