Python-KontAKTOCRScreen

Screens a single PDF document for personal data and suggests what to redact, for the KontAKT aktindsigt (FOI request) system.

KontAKT triggers this once per document after a case's files have been transferred and the caseworker starts OCR-screening. It only suggests — it never changes the document.

What it does

For one document:

1. Downloads the PDF from KontAKT's local file store (GET /api/v1/cases/{id}/documents/{doc_id}/content).
2. Extracts the text — from the PDF's own text layer where it has one, and via Tesseract OCR (Danish + English) for scanned / image-only pages.
3. Scans the text for the personal data that typically must be redacted:
   • CPR numbers — including numbers split across a line break,
   • phone numbers (Danish, +45 or spaced pairs),
   • e-mail addresses. Each hit also records its location(s) on the page — one rectangle per word it covers (so a CPR split across a line gives two boxes) — so the redaction editor can pre-place a box over it.
4. Reports the suggestions back to KontAKT, where a caseworker reviews them.

Names and addresses are deliberately not detected — they're too noisy to be useful as automatic suggestions. The results are heuristic hints for a human, not an authoritative redaction list.

Input (one document)

Field	Meaning
kontakt_case_id	KontAKT case id
doc_id	KontAKT document id (the PDF is read from the store by this id)
source_case_id	GO/Nova case number
dok_id	Source document id

Output

A callback to KontAKT, per document:

{
  "status": "screened",
  "suggestions": [
    {"type": "cpr", "value": "010190-1234", "count": 2, "pages": [1, 3],
     "rects": [
       {"page": 1, "x0": 0.22, "y0": 0.10, "x1": 0.40, "y1": 0.12},
       {"page": 3, "x0": 0.15, "y0": 0.55, "x1": 0.33, "y1": 0.57}
     ]},
    {"type": "telefon", "value": "12 34 56 78", "count": 1, "pages": [2],
     "rects": [{"page": 2, "x0": 0.20, "y0": 0.30, "x1": 0.39, "y1": 0.32}]}
  ],
  "pages": 4,
  "ocr_used": true
}

Rect coordinates are fractions of the page (0–1, top-left origin), so they map onto the page at any render scale.

On failure it reports {"status": "error", "note": "…"}.

Required configuration

• Credential KontAKTAPI — username = base URL, password = API key

Requirements on the worker

• Tesseract (with Danish + English language data) is used to OCR scanned / image-only pages. It is auto-installed if missing (the same way the conversion robots auto-install LibreOffice). Set TESSERACT_PATH to use an existing install instead, or OOMTM_TESSDATA_BASE_URL to change the language-data source (defaults to tessdata_fast).
• If a page can't be OCR'd (Tesseract truly unavailable), that page is left unscreened and the document is reported as incomplete (a ⚠ badge in KontAKT) rather than falsely "clean".

Dependencies

The shared oomtm library (pdf, for the Tesseract auto-install), plus PyMuPDF for text/render and pytesseract for OCR.