Currently, if an index scan is not covering, a fetch is done immediately after the index scan, and filters are evaluated after the fetch. There are situations where a filter cannot be effectively used to generate an index span, however the filter only depends on index keys and thus can potentially be evaluated with the index key values only. Such a filter can potentially discard result returned from the index scan and thus reduce unnecessary fetches.

Such filters can be something like:

lower(c1), when c1 is an index key
c1 LIKE "%abc%", like pattern with leading wild card