When to use vector PDF parsing vs. vision

If you've worked on AI takeoff software you've probably watched a vision model 'OCR' a vector PDF and produce text that's almost-but-not-quite right. The reason is structural: vector PDFs encode every text run, line, and rectangle as geometry. The vision model sees a rasterized image and tries to reverse-engineer that geometry from pixels. Round-trip loss is guaranteed.

The split

• Vector PDFs (Revit, AutoCAD, Bluebeam exports) — use a vector parser
• Scanned plans (rasterized) — use OCR + vision
• Mixed sets — split per-page based on extracted-text length

Our parser is built on pdfplumber. Per page we lift words (with bboxes), lines (with widths), and rectangles. Vision still runs on raster pages — but for vector pages, vector text becomes the source of truth and the vision model is a cross-check, not the primary signal.

This is one of the reasons our cost per bid is lower than the AI-only competition. We're not paying vision tokens to re-derive what the PDF already encoded losslessly.