If the content is to be trustworthy then using LLMs to compress it makes no sense.
Use a universal function approximator to approximate the universe, seek Erf(x)>threshold, interrogate universe for fresh data, retrain new universal approximator, ... loop previous ... , universe in a bottle.