[jlemoine]

Thanks for your feedback, I am the author of the article.

1) You are right that it is possible to mix both popularity and relevance, but you need to use boost and store everything in the float _score. This is dangerous and has side effects (for example you have a big risk of obtaining at some point a hit with typos before an exact one). It is really difficult to control ranking with boosts.

2) The ngrams approach is indeed an alternative. But it also has major drawbacks in term of relevance, mainly for the proximity between terms.

3) Phrase query is a good way to improve performance but it breaks user experience if the terms are not close together (these hits are not in the search results). It's better to let the proximity do its job and impact the ranking.