# Extract Document Data

> Extract structured, grounded fields from documents (payslips, invoices, bank statements, contracts, forms) — values are what the document shows, and missing values abstain instead of being hallucinated. Use when the user asks to pull fields out of a document, convert a document to structured data, or parse invoices/payslips/statements.

- Skill: `sketchjar/extract-document-data-2` (Agent Skill)
- Install (CLI): `npx skillmds@latest add sketchjar/extract-document-data-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/sketchjar/extract-document-data-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Sketchjar (https://skillmd.com/u/sketchjar)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/sketchjar/extract-document-data-2

---


# 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:
     ```bash
     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"
     ```
   - **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.**

   ```json
   {
     "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)
```

## 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