The same sequence can be the result of multiple insertions or deletions. In cases of such equivalent indels, one version may have been called by the variant caller but a different one found ExAC or other databases.
For example, given a reference sequence ATATATATATATATATATATATATATATATATAT. Now suppose an insertion of an additional AT is found in the beginning of the sequence while ExAC has an insertion of an extra AT at the end of that sequence with a frequency of 10%. The two insertions are in fact equivalent because the resulting sequence is the same. Therefore, our system will match the insertion it found to the one in ExAC.
When we inspect the read alignments however, the insertion may not be seen in the same place as shown on our system. Therefore, our algorithm carefully and correctly matches indels to the frequency of their equivalent found in Gnomad and other databases