if our python scripts talked directly to Microsoft, we would be paying $10 per 1,000 requests, even if we had previously OCR'd the same file. It would also take several seconds (because it would process the same file again from scratch).
eh, it's kinda fair when you consider how good it is. It's just stupid that Microsoft doesn't serve you the same result when you make the same request lol. So we built a Worker that fixes that for us.
I know WebSockets are really fast but their stateful objective makes them kind of BS to work with when scaling you can only say handle "regions" and not all users.