Any AI will be flawed, it is human created and fed human content.
Bad content in = bad content out.
Using the internet as a source will only provide generic information with specifics and context being incorrect.
The hardest part of building an agent is sanitising the data sources.
We are also likely to see models retrained on content that contains increasing amounts of already-generated content, so that is the snake eating its own tail.
In Victron’s case you could train your own model, though this is really expensive, or pick one that works and feed it a Victron specific diet of supplementary data that is known to be good.
Feeding it all the (previous) forum topics did not yield good results.
Even then, it will help find better data in a broader context, but, since any AI lacks intuition, when it comes to scenario specific answers and differentiating between relevant scenarios, it still will not be a substitute for an experienced and, ironically, trained, professional.
These aren’t one size fits all systems.
I think we would all be happy if it can bridge the gap for the majority of FAQ’s which represent most of the queries, on this forum at least.