The performance comes from the fact that field like('%search_item%') cannot use any index - it generates a table scan. Note that field like('search_item%') (without the leading "%") is a completely different story - if there is an index for field it will use it - but presumably that's not the result you want.
This is what FTI is for really - it uses the Microsoft full text searching support (you should be able to find the files it builds to support it from the properties of the full text index).
As far as the different results go, I'm not sure - what type of phrases are you looking for? I think the full text search stuff is quite clever about finding approximate matchesm whereas the "LIKE" predicate is a character-for-character match.