Those are simply features in your data for which the database itself does not have sufficient data to accurately assign a taxonomy at that level, but for which the tool can tell that you have enough information in your data to be able to assign to that level, if you had a relevant entry in Silva.
For an example of filtering based on taxonomy, check out this example in the Cancer Microbiome Intervention Tutorial. I do not think you will want to use --p-mode-exact. I think you will simply need to add a trailing semicolon to the query against the taxonomic annotation: `--p-exclude p__;.
As a warning, when you are filtering these out, you are disposing of good reads. Depending on how much data this filtering step eliminates, it might be worth trying to build a more specific or expanded reference database.