01-document-processing/document-structure-analysis/SKILL.md

metadata

name: document-structure-analysis
description: >-
  Map Nordea PDF layout into zones (header, account info, table, footer) and
  compute a layout signature hash for template matching. Use after text
  conversion to classify document regions and detect template version.
metadata:
  category: document_processing
  priority: CRITICAL
  config_id: document_structure_analysis

Document Structure Analysis

Purpose

Zone classification + template fingerprinting. Enables template_library matching (PerfectPDF Skill B) and anchors table extraction coordinates.

Zones (Nordea Tiliote, A4 595x842)

Zone	Y range	Contents
header	15-126	Nordea logo, bank info
account	40-110	Account holder, address, IBAN
period	~130-160	"Kausi: DD.MM.YYYY - DD.MM.YYYY"
table	170-800	Transaction rows
footer	800-820	"Sivu X/Y", disclaimers

Workflow

1. Load spans from pdf-text-conversion
2. Assign each span to a zone by bbox Y
3. Extract anchors: IBAN regex, "Kausi:" period line, "Sivu X/Y" page marker
4. Compute layout signature: SHA1 of (page_count, block_count per page, font set, zone boundaries)
5. Match signature against templates table -> template_id + version

Code

import hashlib, re

IBAN_RE = re.compile(r"FI\d{2}\s?\d{4}\s?\d{4}\s?\d{4}\s?\d{2}")
PERIOD_RE = re.compile(r"Kausi:\s*(\d{2}\.\d{2}\.\d{4})\s*-\s*(\d{2}\.\d{2}\.\d{4})")

def layout_signature(pages: list[dict]) -> str:
    parts = [str(len(pages))]
    for p in pages:
        fonts = sorted({s["font"] for s in p["spans"]})
        parts.append(f"{len(p['spans'])}:{'|'.join(fonts)}")
    return hashlib.sha1(";;".join(parts).encode()).hexdigest()

def find_anchors(full_text: str) -> dict:
    return {
        "iban": (m.group(0) if (m := IBAN_RE.search(full_text)) else None),
        "period": PERIOD_RE.findall(full_text),
    }

Integration

• Feeds: pdf-table-extraction (zone coords), template matching
• Uses: pdf-text-conversion output

Errors & Edge Cases

• Unknown signature -> create new template candidate, flag for human review
• Extra pages (attachments) -> exclude from signature, process separately