You should be able to do a K-means type thing. Where your query is an entire group, and you grab the field from the chunk locally. But you might still be able

You should be able to do a K-means type thing. Where your query is an entire group, and you grab the field from the chunk locally. But you might still be able to use some frequency sampling to predict the words used, unless those chunks are very very carefully constructed.