Extract structured data from documents with grounding

Extract Document Data turns unstructured documents into structured fields with per-value grounding. Each extracted value carries a page reference and a model-reported confidence score; requested fields that are absent are returned in not_found rather than inferred. It is designed for cases where a missing value is safer than a hallucinated one, including lending, accounting, and compliance workflows.

Use this skill to extract fields from payslips, invoices, bank statements, receipts, or contracts, or to convert document contents to JSON for a downstream system. It supports ad-hoc field requests, built-in templates, and schema-free extraction.

Extraction is not document verification. A page citation and confidence score do not prove that a value is correct or that a document is genuine. For authenticity questions, to pair extraction with the verify-document skill first. Reconcile consequential values with the original and authoritative systems before payment, lending, accounting, compliance, or legal action.

The input is a PDF, PNG, JPEG, or DOCX from a URL or local file, plus the chosen extraction mode. Ad-hoc mode names the desired fields; the template mode uses one of the built-in schemas for payslips, tax invoices, bank statements, receipts, or contracts; schema-free mode omits a field list. Tables are extracted with structure preserved, and multi-page documents are processed page by page subject to document page limits.

The response includes the mode, document type, pages read, field values with confidence and page, and a list of requested fields not found. For example, an absent year-to-date tax value belongs in not_found; it must not be converted to zero or guessed. Keep the original file and extraction response so a human can reproduce and correct disputed fields.

A source response example illustrates the structure:

{
  "mode": "schema_free",
  "document_type": "payslip",
  "pages_read": 1,
  "fields": {
    "employer_name": {"value": "Acme Cleaning Pty Ltd", "confidence": 0.95, "page": 1},
    "net_pay": {"value": "2845.10", "confidence": 0.97, "page": 1}
  },
  "not_found": ["ytd_tax"]
}

The skill uses the Stipple API and describes a free anonymous tier, with a weekly allowance and metering at one credit per page read, minimum one. It accepts a built-in template through a form field, and supports table extraction. Invoices, statements, payslips, and contracts can contain personal, financial, or commercial data: obtain explicit approval before uploading to a hosted service, minimize submitted content, and confirm current retention, residency, access, and deletion terms.

Accounting, lending, compliance, and operations teams can use it when document values need to be traceable to pages and absences must remain explicit. A human reviewer should verify consequential values against the original document and authoritative records.