[Kiro]

Why all those steps? Why not just file + prompt to JSON directly?

[tlofreso]

Having the text (for now) is still pretty important for quality output. The vision models are quite good, but not a replacement for a quality OCR step. A combination of Text + Vision is compelling too.