I've played a bit with the new version recently, and got a few comments for it (the old one as well), in no particular order.
I liked the idea of voting for samples, but I think it's also very subjective. There are a few samples that seem perfectly okay for me, but were downvoted. Each person's handwriting is unique, and more so are imprecise scribblings using the mouse. Frankly, I wouldn't like to have to spend too much time perfecting my alphas and omegas -- my handwriting is notoriously unintelligible, as is my mouse drawing. I think voting should only be in the negative direction, if a sample is obviously poor (spam, very bad attempt, or something that doesn't resemble at all the symbol in question).
Offline use is very nice feature, especially if there would be a way to download/update the whole sample database at once and to use it solely in offline mode. This might potentially save some bandwidth from heavy users.
- Integration with TeX IDEs would be definitely be very powerful (i.e. from within the UI, scribble the symbol, click "Submit", and get a list of suggestions), although I don't expect you to do this. Rather, if there's an API is in place that allows communication with the client, third-party developers and power users might step in to provide plugins for their favourite editor1.
- Some symbols might actually be drawn/rendered in very different way, depending on the font you use, and whether you scribble the hand-drawn or the typeset version. Case in point:
\delta. Thus, potentially different classes of samples should be accepted for some (or even all) symbols.
- Selecting symbol class might also go a long way towards accurate recognition. Most of the times, one would know what kind of symbol they need -- a foreign letter, mathematical operator, dingbat, relational symbol, glyph "beautifiers" (diacritics, for example), etc. So, if my symbol is pretty simple in shape, chances are I will end up with a lot of false positives that are not interesting.
- Already mentioned diacritics in the previous point, and although they are not all that many, maybe they could also be recognized by detexify. For example, I always forget how do I specify breve accent
\u{}, and have to disrupt my writing to check somewhere.
- On the character description, some indication of whether it's available in math or text mode would be helpful.
- Similar or counterpart symbols (mostly applicable in math mode) might also get preferential treatment. For example, if I draw the parallel sign "||", showing explicitly "not parallel", even if it's not with top score, would be useful. In the previous detexify version, the "normal" symbols were recognized easier than the negated ones (at least in my experience).
- Major Feature: Sometimes I would see unknown symbol in written text or online document. Instead of sketching it, perhaps it would be a better idea to just provide a contrast image of it, and let detexify work its magic on it, instead of me trying to sketch it much less precisely by hand. Bonus points for uploading an image with text and making a box around the symbol that interests me instead of doing the tedious pasting-panning-resizing-saving-uploading on the desktop.
Apart from that, the sample database is currently tiny, but I don't hold it against you -- it would be a significant effort by itself to have a rich database with quality samples.
Although this seems like an enormous list of suggestions and complaints, I can assure you that detexify is extremely helpful for a lot of people (myself included), and we certainly appreciate your effort to provide a useful service.
Good luck!
1 I wonder how would that work for vi/Emacs users...
\alpha,\betaand\gamma. Thanks for the work, though!