It seems I can use the combo of a shingles filter and tokenizer to create a multi-word (phrase) autocomplete feature using a multi-match query. However it also looks like I can use the Completion suggester for basically the same purpose. If I have mostly 2,3 or 4 word phrases that make up the suggestions, how big of an index can I make before the Completion suggester performance degrades. It seems that is the whole purpose behind the Completions suggester - speed.
I plan to use a separate index for the autocomplete feature.