Split Classification And Segmentation

A single uploaded PDF is actually a bundle: a cover letter, an invoice, and a W-9 concatenated. The integrator calls /split to segment it into constituent documents but assumes a fixed three-way split for every bundle.

Use /split to detect document boundaries within the bundle and return per-segment page ranges + classifications, treating the segment count as data-driven (not a fixed assumption). Carry each segment's page range forward so downstream /parse or /extract operates on the correct sub-document. Verify …

Integrator configures /split with a set of category labels for classification but provides vague, overlapping labels ('document', 'paperwork', 'other').

Provide a mutually-exclusive, well-described label set aligned to downstream routing (e.g., 'invoice', 'contract', 'tax_form'), with an explicit 'unknown' bucket for low-confidence cases. Overlapping labels make classification non-deterministic and unroutable. Measure per-label precision/recall on …

A scanned bundle includes some pages rotated 90/180 degrees (fed sideways through a scanner). The integrator assumes all pages are upright and segmentation/classification degrades.

Rely on Reducto's orientation handling (de-skew / rotation detection) rather than assuming upright input, and spot-check segmentation on rotated samples. If orientation handling is not documented as automatic, pre-normalize orientation before submit. Treat automatic de-skew/rotation as [REQUIRES-VE…