Measured fidelity first
Sample the corpus and verify extracted text against the rendered page. Never trust a text layer from the file extension alone.
Premium skill
Gets messy sources into faithful structured text: sampled fidelity audits, parsing strategy per document class, lossless tables, a normalized contract with provenance, and golden-sample fidelity evals.
$ npx skills add β¦ --skill ingest-messy-documentsSign in with Premium to generate a fresh install link.
Browse the authored file tree β tools, references, and evals included. Source unlocks with Premium.
SKILL.mdmarkdownBrowse every authored SKILL.md with premium.
File tree for ingest-messy-documents stays visible so you can see the pack before unlocking.
Sample the corpus and verify extracted text against the rendered page. Never trust a text layer from the file extension alone.
Choose text layer, OCR, vision transcription, or hybrid from sample evidence, with cost and fidelity tradeoffs stated.
Preserve reading order, headings, and tables as structure with cell provenance. Never flow grids into prose.
Emit normalized documents with stable ids, content hashes, and citation provenance for Ground and Extract.
Natural partner for Ground AI SDK Agent (retrieval after readable text) and Extract Structured Data (typed records from the normalized contract). Fidelity evals follow Build AI SDK Agent Evals conventions.
A RAG chatbot that was confidently wrong because PDF text layers were empty on scans.
A table-heavy policy corpus where OCR scrambled cell values until a lossless table path shipped.
Feeding Ground so citations still open the source page after messy PDFs are normalized.
Monthly document updates where unchanged files skip re-parse and deletes finally leave the index.
Available with Premium. The private install link expires; the downloaded skill does not. Generate a fresh link when you want to update it.
Lifecycle skills and UI craft skills