Extract structured data from documents with grounding

This skill extracts structured JSON from documents - payslips, invoices, bank statements, receipts, contracts - with per-value grounding: every extracted value cites the page it came from and a confidence score, and any value that isn't clearly present is reported in a not_found list rather than invented.

Use it for parsing payslips, invoices, bank statements, receipts, or contracts into structured JSON for downstream systems, especially anywhere a hallucinated value is worse than a missing one - lending, accounting, compliance. This is extraction, not verification: confidence and page grounding do not prove a value is correct or that the source document is authentic. Pair it with the verify-document skill first when genuineness matters, and reconcile consequential values against the original document and authoritative systems before payment, lending, accounting, or compliance action.

Input is a document by URL or local file path (PDF, PNG, JPEG, DOCX) plus an extraction mode: ad-hoc named fields, a built-in template (payslip, tax_invoice, bank_statement, receipt, contract), or schema-free.

curl -X POST https://www.stipple.sh/v1/extract \
  -F "file=@payslip.pdf" \
  -F 'fields=[{"name":"employer_name"},{"name":"net_pay"},{"name":"pay_date"}]' \
  -H "Authorization: Bearer $STIPPLE_API_KEY"

Output is a JSON object per requested field with its value, confidence score, and source page number - for a payslip, that might read employer name at confidence 0.95 on page 1, and net pay at confidence 0.97 on the same page - plus a not_found list for anything requested but not located, and a pages_read count. Absences are reported, never guessed: a missing year-to-date tax figure comes back as "not found," never as zero.

The Stipple API (free anonymous tier, or a free key for higher metering) performs the extraction, at a cost of 1 credit per page read, minimum 1 per document. Tables are extracted with their structure preserved, and multi-page documents are processed page by page. It pairs naturally with verify-document, run first, since an extracted value from a tampered document is still wrong even when correctly read off the page.

Teams converting unstructured documents into structured data for lending, accounting, compliance, or other downstream systems, where a wrong or invented value is more dangerous than an honestly reported gap and a human reviewer needs to be able to trace every value back to the exact page it came from before anyone acts on it, especially when the same document is later reconciled against a source system.