The question.
Document extraction fails when a plausible field value loses its page position, table context, or connection to the original evidence.
The architecture combines layout-aware extraction with schema validation and a review queue for uncertain or inconsistent fields.
The proposed approach.
- Parse text, layout, tables, and image regions while preserving page coordinates
- Extract into explicit typed schemas with field-level source references
- Check totals, required fields, and cross-document consistency
- Send uncertain outputs to a reviewer with the relevant source region
Target architecture.
PDF / imageLayout and OCRStructured extractionValidationEvidence-linked reviewExport
The difficult parts.
- Complex tables and scanned documents
- Code-switching and numeric formatting
- Distinguishing a missing value from a guessed one
How to evaluate it.
- Field and table accuracy by document family
- Source-coordinate correctness and confidence calibration
- Review effort and correction rate
Planned deliverables.
- A document and extraction schema collection
- An evidence-linked review workflow
- A document-family evaluation protocol
Design principles.
A missing value is better than a confident invention.
Keep extracted facts connected to their original page.