How It Works

OCR engines return an unordered bag of bounding boxes: [x, y, width, height, text]. Reading them in naive stream order scrambles columns, detaches prices from titles, and cuts descriptions in half at page boundaries.

afferent recovers the document layout in four passes:

1. Split Columns

Horizontal projection histograms detect empty gutters across the page. This isolates multi-column text bands from full-width spanning banners, section headers, and footers.

• Text from Column A never bleeds horizontally into Column B.
• Spanning elements (banners, disclaimers) are extracted into spanningLines.

2. Lock Baselines

Titles and labels are paired with right-aligned values (prices, SKUs, dates) sharing the same visual horizontal baseline (|Δy| <= 0.012 by default).

• Handles wide whitespace gaps between label and value.
• Supports vertical layouts where a price sits directly below a title.
• Recovers same-line secondary attributes (e.g. per-unit suffixes).

3. Stitch Blocks

Within column boundaries, wrapped titles, narrative paragraphs, and sub-bullets merge into unified item blocks:

• Merges multi-line descriptions below a title based on vertical proximity.
• Recovers stub titles and splits colon-delimited narrative headers.
• Computes the exact mathematical union bounding box covering all constituent lines.

4. Bridge Page Breaks

When an item begins at the bottom of Page N and concludes at the top of Page N+1, afferent evaluates:

• Punctuation (dangling commas, missing terminal periods).
• Grammar & casing (lowercase starts on the new page, trailing conjunctions).
• Value complementarity (title on Page N, price on Page N+1).

Unambiguous continuations stitch automatically. Ambiguous boundaries can query a System 1 model (such as Jev) for semantic confirmation.