Right now I’m working on creating some special materials to add some depth to the crafting system in-game. One of my players is a forge cleric (a priest who venerates a deity associated with crafting, specifically metalwork), so he’s been champing at the bit for a better crafting system.
I’ve got my Python app building a conversation to feed the GPT-4 model with all of the information needed (resource types, subtypes, rarity odds, what JSON schema needs to be returned, acceptable values for each property, etc). Additionally, I’ve found I need to feed the API back the history of what it’s already generated or it frequently generates repeats (materials with the same name and/or basically the same uses). However, feeding the history back to the api rapidly increases the tokens used in each message. Currently I’m feeding back a summary of the history (names and descriptions of materials only), and that seems to be helping somewhat.