Not crazy at all. If it does what you say, then it is very useful.

But "trust me" is not enough -- you have to prove it works in a wider context. IR has many pitfalls, and success in one domain is not an indicator of general applicability. Your algo may work well on the subset of Wikipedia pages. It may spit out garbage with other sources or on search terms that would not occur to you.

Just as in crypto, IR algorithms have to be kicked around a lot before they get serious attention. I suggest you put something up that people can play with. You don't have to reveal the algo.