I found different solutions to do it and I finally decided to merge them in a single "optimal" code.
During the typing it checks each new word (from the previous whitespace to the current pointer) and proposes to complete it with words included in a array. Predicted words are shown in a list under the input field and can be selected with mouse or with up/down/enter keys.
These are the sources considered to create this code:
'?do=embed' frameborder='0' data-embedContent>>