zirpu

Fork

Use FAISS as a basic database of thoughts.

IRC stuff.

Multiple channels, timestamps, better formatting of log lines, log
storage, and small tweaks.

Refactor IRC simulation.

Now we're connecting to a real server!

Wire up llama.cpp to the simulated IRC channel.

It all works. llama.cpp is slow for reasons I don't fully understand,
but it will get faster soon. In the meantime, this is good enough to
move forward with experiments that use a hybrid of HF and llama.cpp
agents in a single simulation.

Unify HF text-generation utilities.

The idea will be to introduce a companion module which implements a
llama.cpp loader.

A simulated IRC room.

This is too much overhead and I'm going to refactor it into IRC bots.

Cut off the log with a fast bisection search.

This results in fewer calls to the tokenizer and also allows for the
amount of the log included in the context to grow if the prompt happens
to shrink.

Factor out another few classes.

Factor out Timer class.

Refactor prompt logic.

Now the default action is to ask for another continuation, there's no
magic values -1 and 0, reprompts can be continued multiple times by
default... It was worth thinking about for half a week.

Add shadow selves.

Add two more utilities.

Lot of duplicated code, but not much code total. Suggests that we should
combine these into a single script with subcommands.

Accept some last-minute personality traits.

Akin to the scenario, this allows for gentle steering of a conversation.

Start recording progress.

This is a simple script for character conversations, in the tradition of
AIDungeon or CharacterAI. I don't bother trying to parse English or
synthesize intermediate sentences; contrast with AIDungeon to be
enlightened.