Accounting Skill

Extract structured data from accounting documents (invoices, POs, bank statements) into Excel tracking sheets with JSON backups. Handles digital PDFs, scanned PDFs, and images via automatic OCR.

Prerequisites

Install system OCR dependencies before first use. See {baseDir}/references/ocr-setup.md for full guide.

# Ubuntu / Debian
sudo apt install tesseract-ocr tesseract-ocr-vie poppler-utils

# Verify
uv run {baseDir}/scripts/ocr_utils.py check

Quick Start

1. Classify an unknown document

uv run {baseDir}/scripts/classify_document.py /path/to/document.pdf

Returns JSON with type (invoice / po / statement / other), confidence, and a ready-to-run extraction command.

2. Extract an invoice

uv run {baseDir}/scripts/extract_invoice.py /path/to/invoice.pdf -o invoice_tracking.xlsx

Appends to the Excel tracking sheet. Use --dry-run to preview parsed data without writing.

3. Extract a bank statement

uv run {baseDir}/scripts/extract_statement.py /path/to/statement.pdf

Creates statement_{bank}_{date}.xlsx with transactions. Use -o to specify output path.

4. Extract a purchase order

uv run {baseDir}/scripts/extract_po.py /path/to/po.pdf -o po_tracking.xlsx

Tracks delivery dates and flags overdue/urgent POs.

5. Generate empty Excel templates

uv run {baseDir}/scripts/generate_templates.py all -o ~/accounting/

Creates blank tracking sheets: invoice_tracking.xlsx, po_tracking.xlsx, statement_template.xlsx.

Common Options (all extractors)

Workflow

Single Document

File → classify_document.py → route → extract_*.py → Excel + JSON

Batch Processing

For a folder of mixed documents, classify first, then route:

for f in /path/to/docs/*; do
  uv run {baseDir}/scripts/classify_document.py "$f" --output-dir ~/accounting/
done

Then run the suggested extraction commands from each classification result.

OCR Strategy

All scripts share {baseDir}/scripts/ocr_utils.py which auto-selects the best extraction method:

1. Digital PDFs → pdfplumber (fast, no OCR needed)
2. Scanned PDFs → pdf2image + pytesseract at 300 DPI (fallback when pdfplumber gets <50 chars/page)
3. Images (JPG/PNG/TIFF) → pytesseract with grayscale preprocessing

Each result includes ocr_confidence and extraction_confidence percentages. Documents below 85% are automatically flagged needs_review.

Validation Rules

• Invoices: Subtotal + VAT = Total (auto-checks math), duplicate detection by invoice number + vendor
• Bank statements: Opening balance + credits − debits = closing balance
• POs: Delivery date tracking with overdue/urgent alerts

Reference Documents

Read these for field schemas, Vietnamese format details, and validation logic:

• {baseDir}/references/invoice-fields.md — Vietnamese VAT invoice fields, tax rates, patterns
• {baseDir}/references/bank-formats.md — Vietnamese bank names, transaction formats, amount patterns
• {baseDir}/references/po-fields.md — PO fields, delivery status logic, payment terms
• {baseDir}/references/ocr-setup.md — OCR installation, troubleshooting, confidence scoring