Turns out sometimes docid can be a string.

For small counts, this reported wrong number of documents (batch size
instead of the small count).

This allows to query large sets in batches and send the results
upstream.