Usage docs

Detection

How CommonForms detection works, how confidence levels are used, and what to adjust when candidates look wrong.

How to use this docs page

This page is meant to answer one operational stage of the DullyPDF workflow well enough that you can run a controlled test without guessing. Read the sections below, validate the behavior against one representative document, and only then move to the next linked page.

That order matters because most setup failures come from mixing detection, mapping, fill validation, and sharing into one unstructured pass. A narrower review loop keeps troubleshooting faster and makes the template easier to trust once you save it for reuse.

What detection returns

Detection returns a field list with key values: name, type, page, geometry (rect), and confidence metadata.

Geometry is normalized to top-left origin coordinates and rendered as {x, y, width, height}in the editor.

Field types supported in the UI are text, date, signature, checkbox, and radio.

Confidence review

  • High: confidence >= 0.60
  • Medium: confidence >= 0.30 and < 0.60
  • Low: confidence < 0.30
  • Numeric confidence parser accepts either 0..1 values or 0..100 percentages (for example 82becomes 0.82).
  • Start review from low-confidence candidates because they drive most downstream errors.

Common limitations and fixes

  • Low-quality scans can reduce field boundary precision.
  • Dense pages may produce close candidates that need manual cleanup.
  • Decorative boxes can be mistaken for fields; remove or repurpose them in inspector.
  • Encrypted PDFs are rejected and must be unlocked before detection.

PDF quality rubric

  • Best: native PDFs with high contrast, clear form lines, and predictable spacing.
  • Usable with review: scans that are readable but have light skew, compression noise, or inconsistent line weight.
  • High-risk: faint scans, dense tables, decorative borders, or layouts where fields are packed tightly together.
  • The dirtier the PDF, the more important it is to review low-confidence candidates before rename or mapping.

When to redraw instead of resize

  • Resize when the candidate is fundamentally the right field but the geometry is slightly off.
  • Redraw when a decorative box was mistaken for a field or when the detection captures the wrong label/line pair entirely.
  • Delete and recreate when the current candidate would require several compensating edits that are harder to audit later.

Geometry values and editor constraints

  • Rectangles are clamped to page bounds during drag/resize.
  • Minimum field geometry is type-based: text/date/checkbox = 12 points, signature = 16 points.
  • All geometry edits in inspector and overlay are applied in the same coordinate system.