Sphinx's proximity-enabled ranker uses a slightly modified BM25 ranker (statistical bag-of-words) + a longest-word-substring match formula strongly favoring the latter, while Solr uses some other statistical ranking function (not BM25, but similar) + a boost if desired for word bigrams (this is similar to the LWS approach). I think both of these don't model a human's view of relevance, whereby relevance doesn't fall off a cliff when words in an answer aren't necessarily adjacent or in the same order.
Simple examples:
Query: Bob Jones
Body: . . . . Jones, Bob . . . . (looks relevant to me, but this will fall back to statistical-only)
-or-
Body: . . . . Bob MiddleName Jones . . . . (same)
I know there is a cost to this, but I can't be the only one who noticed that essentially both Solr and Sphinx will fall back to the bag-of-words statistical ranker if the words are out of order or separated by a word, which could even be a stop word in some cases.
Thoughts? What if I want to rank either of the cases above higher than those where the words just appear somewhere in the document? Or am I wrong and does Solr or Sphinx do this?
