Major Update

I've been working on-and-off on this update for quite a while, as various changes to the dependencies (the LLM and Gelbooru's image servers, among others) have stopped me from releasing intermediate versions. 

Critical changes:

• Gelbooru has changed image hosting, which broke the image engine. This is fixed now, and loading should work much more consistently to boot.
• Cohere released a new LLM, Command-A, to supercede Command-R+. While integrating it was a huge hassle, as it was broken in the API for about a month, it is now working, providing a much stronger LLM narrator with better creative writing and reasoning skills.

Other Changes:

• Total overhaul of the LLM backend: the LLM is now prompted with a more complex reasoning process, improving the narration and making the narrator more firm in reining in the player when necessary, making cheesing the narrator less tempting. Further,  the switch to Command-A finally allows for a longer chat history to be maintained (optional--it slows the API calls down). To allow the model to focus on the reasoning tasks, many tasks previously handled by the LLM have been offloaded to the local embedding model.
• Dynamic reaction images. Unlike Sketchy Mind Control App, where these images were AI generated and selected by the LLM, I've greatly streamlined this feature to make it scaleable: images are sourced from Gelbooru, as with the narration images, preprocessed to find faces, and the image is loaded and dynamically cropped to the character's face at runtime. They're also selected using the embedding model--comparing the narration to predetermined descriptions of each expression. This takes the load off the LLM and still yields excellent results. This approach does mean that there are fewer expressions than in SMCA, however, as not all expressions have common gelbooru tags identify them.
  
• Many tasks are now handled by the embedding model by comparing a variety of embedded text descriptions of key gamestate changes to the narration text. This is vastly more efficient that asking an LLM to assess them, and many state changes can be identified this way as well or even better than as part of a multi-part structured query. For subtle state changes, the embedding queries are run in tandem with the LLM queries, giving the game two chances to make the correct state change and improving consistency. You should see more consistent NPC travel, NPC focus shifts, stamina changes, etc. This also means multiple npcs can travel simultaneously.
• A "twist" mechanic has been added. Every 10 turns, the narrator will generate a new twist to the story, adding more variety between playthroughs and avoiding a repetitive experience. These twists are intended to be small and to be woven organically into the story rather than forcing a sudden pivot, but I'd love feedback from players as, by its nature, the outcomes are highly variable and hard to thoroughly playtest.
• A third method of generating characters has been added. Previously, characters were premade (default or community created) or procedurally generated, with the LLM only generating a bio for the randomized character sheet. Now, I've added the ability for the narrator to add new characters as needed. In this method, the narrator provides a description of the character, and, using the local embedding model, that description is mined for details to create a full character sheet. This allows characters to be added to meet specific story requirements, rather than forcing the narrator to shoehorn story beats onto a procedurally generated character whose character sheet is a poor fit for them.
• To make the above feasible, proper GPU acceleration has been added for the local embedding model, making it about 10-100x faster with appropriate hardware. Even integrated GPUs will benefit greatly.
• Lots of QoL improvments: Better handling of travel, improved bio generation for procedurally generated NPCs, and more.