[PaulHoule]

I know how to make some ranking functions that are good in some ways. They are not good in other ways so I don't break Arrow.

The shadow of Arrow however is that it is a lot of work to get a real improvement over a tfidf with random characteristics and you are limited by an asymptote which is less far away than you'd like: the rel teams at Google and Bing cannot beat the p=0.7@1 barrier for text search because 0.3 of the time they will get your intent wrong.

There are ranking functions in Lucene that can be tuned up as I described by making a test data set and using the methods developed at the TREC conference. That kind of tuning really works. At that conference you might see people fight over a point or two of AUC and I don't know if users can feel that but I am sure users can feel the 15 point and more difference we were seeing from tuning.

The odd thing is that hardly anybody does it and the knowledge seems pretty obscure. (e.g. I have talked to people at Lucene, OpenText and other full text search vendors and they are much more impressed with having 10,000 connectors than with the search results being good.)