From what I'm reading, it doesn't really seem like it's distillation at all: they have vast quantities of unlabeled (query, document, <engagement data>) tuples, which we usually, in search relevance, generate labeled tuples directly from, via some heuristic (e.g. clicks).
(cont)
