What problem does it solve? Standard tools like pdftotext corrupt subscripts, superscripts, Greek letters, and square roots when extracting text from technical PDFs, producing equations that look correct but are wrong. This Skill converts PDFs (technical standards, papers, manuals, reports) through a local docling service that preserves equations as LaTeX, extracts figures, and records provenance metadata so extracted content can be cited with confidence. ## Core Features & Use Cases - Faithful PDF-to-Markdown conversion: Uses the local docling service with formula enrichment to deliver equations in LaTeX, plus figure extraction via format=zip so image links never break. - Provenance and verification workflow: Writes a provenance header (source, date, flags, losses) and guides a cheap-to-expensive verification order: recalculate structured data, crop regions at 300 dpi, read full pages only when needed, delegate bulk reading to subagents. - Pitfall handling built in: Decrypts encrypted PDFs with pikepdf before upload, keeps LLM enhancement off to avoid silent factual corruption, and selects a figure-description VLM that fits in 8 GB of VRAM. - Use Case: Convert a 124-page scanned engineering standard into Markdown, extract key pages as PNGs, recalculate a table formula to catch a row-shift defect, and produce a verified transcription file alongside the full extraction. ## Quick Start Convert the PDF standard at ~/refs/NORMA.pdf into Markdown with provenance, extracting pages 12, 31, and 57 as images for verification, using the docling-local skill.
