LDA models need on-the-spot inference capability
The current iteration of on-the-spot paper processing is written around the assumption that the topic model is an sklearn.decomposition.NMF
instance, because that's what has worked best for us so far. However, given the recent success of Gensim models within our codebase, we really ought to add a codepath that supports performing online inference (and potentially online training?) for those as well.