Comment by semiquaver

> Its about being able to answer questions thats not in its training data.

This happens all the time via RAG. The model “knows” certain things via its weights, but it can also inject much more concrete post-training data into its context window via RAG (e.g. web searches for documentation), from which it can usefully answer questions about information that may be “not in its training data”.