adu-pdf-extraction

Extracting text from construction PDF plans requires a vision-first approach because standard extractors cannot parse composited multi-page layouts, diagonal watermarks, or CAD-rendered text. Claude Vision reliably handles drawing pages and rasterized reports by generating page PNGs and per-page markdown for agent navigation.

Structuring an ADU permit plan binder involves generating a JSON manifest with per-page metadata, topic keywords, and spatial drawing zones. This allows agents to route corrections letters to exact sheets and drawing zones without loading the entire multi-page construction PDF into context.

Processing California ADU permit PDFs requires Tesseract OCR and ImageMagick alongside pdftoppm. These dependencies convert binder pages into PNGs and cross-reference OCR text with vision extraction to prevent hallucinations on critical numeric values like dimensions and PSI.

OCR cross-referencing prevents vision hallucinations by combining Claude Vision extraction with Tesseract OCR text validation. This hybrid numeric accuracy approach ensures critical values like dimensions, PSI, and percentages extracted from construction plan binders are verified against the rasterized text layer.

Matching city corrections letters to specific architectural sheets uses the structured JSON binder manifest containing per-page metadata and spatial drawing zones. Agents query this manifest to route each correction item directly to the exact sheet and specification needed without loading every page.

pdftotext returns missing text from Title 24 energy reports because these documents are rasterized images with no selectable text layer. The adu-pdf-extraction skill uses Claude Vision as the primary extraction method to process these rasterized reports where standard text extractors fail.