Tags are sorted in the wrong direction, DESC instead of ASC, and differentiates capitalization.

I don't know how other web applications process tags, but my intuition at this point is that we would get the most mileage out of treating the tag field as a string of full text, and building up a tag index by parsing the text (using all clues, whitespace and punctuation), and then further reducing words that are not proper nouns to their morphological stem (so a tag like "parsing" would be indexed as "parse").

Phrases then just become sequences of tags. If you search for the tag-pair "java compiler" projects that contain both tags in that order and side-by-side would score highest, projects that both tags elsewhere would score less high, and projects with only one of the two tags would score lowest.