The algorithm works well but it's not perfect, so there will be some mistakes. If you really like a haiku, double check the syllable count with your fingers just to make sure.
The precise option stops the program from guessing the amount of syllables if it isn't in the database.
The other two options are to limit the results to those that might actually make some sense. The names are self explanatory. "Start with capital letter" only displays those haiku whose first letter is capitalized. "End with full stop" only displays those haiku whose last character is either a period, question mark, or exclamation point. With both of them checked, the results are usually intelligible.
Please only put in files that display text or html. I would rather not have it search through an MP3 or Jpeg for haikus. If the size of the file is quite large, it will take some time to search through the whole thing, so please be patient. With (S) and (E) unchecked you'll be surprised at how many will be returned.
The system does not handle acronyms or numbers well at all. It assumes they're all words. For acronyms where the letters are pronouced as a word, like RAM, DIMM, NASA, and NAMBLA, that's fine. But when the letters are said one-by-one like in ABC, USA, PC, and AMD it simply doesn't work. Most abrieviations are a problem as well.
It used to be that all numbers were deleted from the text and ignored completely. Now, at least numbers 0 to 20 are converted to their textual counterparts. Any number over that, however, is still ignored.
One more thing: I've put a 2MB limit on the size of files you can input. Large files will take some time to download and process.
UPDATE (Feb 1, 2011)
OMG, I made it way faster. It will still take some time, though, so be patient.