This is an issue for non-latin wikis, the default asciifolding we use will only fold latin (ascii) chars. For some language such as greek it is useless.
Some language specific analyzers may provide better folding but those are not used by autocomplete/nearmatch searches.
We cannot integrate it as is because it lacks the preserve_original option. This option allows the filter to emit the unmodified token at the some position allowing more precise searches if the query includes diacritics.
(currently enabled on greek wikipedia)