I'm interested in the same thing. Some things I've found:
OCRTerminal bills itself as a "Free web OCR service", but the account is limited to 30 docs. And the recognition on my iphone page image was completely useless.
I really thought that anything could beat OCRTerminal's performance, but I when I tried Qipit, I couldn't get it to even show me that I had uploaded a document. Maybe it really wants me to email to it, instead of using their webform. Meh.
Evernote does pretty well searching for keywords in your images, but I haven't found a way to get the text out of the service.
For engines to use to /make/ such a web service:
http://groundstate.ca/ocr Has an 5/2007 review of several open source linux OCR packages.
Tesseract or OCRopus look like probably the best bets.
Cuneiform is supposed to be an enterprise-ready, internationalized open source OCR system, but the web pages I've found seem very incomplete and/or poorly translated: http://www.cuneiform.ru http://en.openocr.org Someone ported the source to linux: https://launchpad.net/cuneiform-linux