Ocr And Layout Fidelity

An OCR'd handwritten form yields plausible-but-wrong characters in a numeric field (a handwritten 7 read as 1). The integrator persists it with no validation.

For high-stakes numeric/identifier fields, validate OCR output against domain constraints (checksum, length, format regex) and route failures to review. OCR on handwriting is error-prone — never persist a raw OCR numeric to an irreversible action without a validation gate. Track per-field OCR error…

Each OCR'd block carries a bounding box. The integrator wants to highlight the source region in a viewer but assumes pixel coordinates when the API returns normalized (0-1) coordinates.

Confirm the documented coordinate space (absolute pixels vs normalized 0-1 vs PDF points) and the origin convention before rendering highlights, and validate on one document that a known block highlights correctly. A coordinate-space mismatch silently misplaces every highlight. Treat the coordinate…

Justified text hyphenates words across line breaks ('inter-' then 'national' on the next line). The integrator concatenates lines verbatim and ends up with broken tokens that hurt embedding and search.

Rely on Reducto's text reconstruction to re-join hyphenated words and normalize intra-paragraph line breaks into flowing text, rather than naive line concatenation. Verify on a justified-text sample that hyphenated words are rejoined. Broken tokens silently degrade retrieval recall.