Thank you for the suggestion. I actually did try restricting the terms by measuring correlation between columns -- the idea being that more discriminating term

Thank you for the suggestion. I actually did try restricting the terms by measuring correlation between columns -- the idea being that more discriminating terms should be searched first. This did result in modest speedups. Fortunately or unfortunately, we were already achieving pretty good speed with Elasticsearch so we didn't implement it. However, it didn't occur to me to try a MoreLikeThis query, which should be even simpler -- I will look into it!