One step at a time :)

What you suggest would benefit from the performance/feedback from what the OP suggests.

This is basically what I am interested on working on currently , using "clues" gleaned from what the user is doing (a la RescueTime) to reorganize/reformulate a user's search query.

First reason is to minimize error rather compound it. Second reason is to avoid overloading the user with information. I do expect that the set of queries prepared for a user will be large and possibly imperfectly ranked, at least initially. To use this set in its raw form to pre-fetch answers will dilute accuracy and be a large amount of information.

Rather, I take the approach that search is a commodity that is already fast enough and good enough. I let the user choose (and thereby rate/vote up) a query and then fetch only relevant results.