extract-document-data – skill Tài liệu & Office cho agent AI

Extract Document Data

Extract structured JSON from documents with per-value grounding: every extracted value cites where it came from (page number, confidence), and values that aren't clearly present are reported in not_found rather than hallucinated. Uses the Stipple API (free anonymous tier).

When to use

• Parsing payslips, invoices, bank statements, receipts, or contracts
• Converting unstructured documents to JSON for downstream systems
• Any extraction where hallucinated values are worse than missing values (lending, accounting, compliance)

Instructions

1. Get the document. URL or local file path (PDF, PNG, JPEG, DOCX).

2. Choose the extraction mode:

• Ad-hoc fields — tell the API exactly which fields you want:

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

• Template — use a built-in schema: payslip, tax_invoice, bank_statement, receipt, contract
• Schema-free — omit fields and let the model extract what it finds

3. Interpret the response.

   {
     "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"]
   }
   

• Every value carries confidence (the model's self-report) and page (grounding)
• not_found[] lists requested fields the model couldn't find — absences are reported, never guessed
• pages_read shows how many pages were processed (page limits apply per document)

4. Report honestly. This is *extraction, not verification* — values are what the document shows, not proof it's genuine:

• "Employer: Acme Cleaning Pty Ltd (confidence 0.95, page 1)"
• "ytd_tax: not found in document" — never "ytd_tax: 0" or a guess
• For "is this document genuine?", pair with the verify-document skill first

Output format

Payslip fields (grounded, not guessed):

  Employer          Acme Cleaning Pty Ltd  (confidence 0.95, page 1)
  Employee          J. Citizen             (confidence 0.98, page 1)
  Net pay           2,845.10               (confidence 0.97, page 1)
  Superannuation    268.20                 (confidence 0.93, page 1)

not_found: ytd_tax
(absences are reported, never hallucinated)

Limitations and Safety

• Invoices, statements, payslips, and contracts often contain sensitive personal,

financial, or commercial data. Obtain explicit approval before uploading them to

a hosted third party, minimize the submitted content, and confirm current

retention, residency, access, and deletion terms.

• Confidence and page grounding do not prove that an extracted value is correct or

that the source document is authentic. Reconcile consequential values against the

original document and authoritative systems before payment, lending, accounting,

compliance, or legal action.

• Keep the original file and extraction response so a human reviewer can reproduce

and correct disputed fields.

Notes

• Costs 1 credit per page read by the model (minimum 1); free weekly allowance applies
• Templates: payslip, tax_invoice, bank_statement, receipt, contract — pass as the template form field
• Tables are extracted with structure preserved; multi-page documents are processed page by page
• Pairs with verify-document (run first, for authenticity) — an extracted value from a tampered document is still wrong
• Free key at https://www.stipple.sh for metering beyond the anonymous allowance