Hi @Circle,
It depends on what you're hoping to do with the representative sequences.
If you want to do fragment insertion, then yes, that would be the appropriate set of sequences.
I'm not sure its the best choice for other applications. The output of reconstruct-fragment-rep-seqs provides consensus sequences only if a reference sequence can't be resolved (ids that have a | symbol in them.) It's designed specifically to integrate with the fragment insertion and makes some downstream assumptions, including that you have the same database and insertion tree version.
The rep-seqs wouldn't be appropriate for use with a naive bayseian classifier, though. (You should do taxonomic reconstruction through sidle.)
If you're thinking about blasting sequences, I think I'd use the original sequence representative sequences from the database as my representative sequence, and then blast both sequences in an unresolved sequence.
Best,
Justine