Ross Intelligence was more a search interface with natural language and, probably, vector based similarity. So I suspect they were hosting and using the corpus in production, not just training a model on it.