@simon the "Limit additional context in retrieval-augmented generation (RAG)" and "Hidden reasoning tokens" seem to be contradictory, or at least at odds: reduce the context, but you cannot use any reasoning tokens that came up while deciding what was and wasn't relevant.
I've been trying to find ways to improve rag index vectors, effectively trying to extract some internal state from a model as a sort-of pre-compiled version of the text - and these hidden reasoning tokens seem to be exactly related to what I'm looking for, but nope we're not allowed to see'm !