> I considered loading another wiki of something different, but it's less convincing as a proof of concept because it needs to allow addition, change and exclusion of small bits of data without the bot going offline for retraining 3 times a day.

I'm not sure how privateGPT works, but any sane implementation is going to store the data in a vector database, not constantly train it into the llm itself. Updating it is easy, fast, and can probably be done without going offline.

...But yeah, the key is somehow replicating your wiki closely enough for a demo. With access to the wiki, maybe you could copy the structure (but not any data) and populate it with dummy data from an LLM? If they are good at anything, its producing something convincing and totally made up.

EDIT: Also, if the backend is llama.cpp, recent advances will make it very easy to run, but questionable for commercial use depending on the model it uses. LLaMA itself is basically illegal to use, even if Facebook doesn't seem to care.