Invoice Scan

Invoice Scan

Extract structured data from any invoice — printed, handwritten, multi-language, stamped.

Setup

Install dependencies once:

cd {SKILL_DIR}/scripts && npm install --production

Set API key — check env ANTHROPIC_API_KEY. If not set, ask the user:

export ANTHROPIC_API_KEY="sk-ant-..."

Scanning

CLI

node {SKILL_DIR}/scripts/cli.js scan <image-or-pdf> [options]

Options:

• --format json|csv|excel (default: json)
• --output <path> — write to file (xlsx for excel)
• --api-key <key> — override env key
• --accept strict|relaxed|any — doc type filter (default: relaxed)
• --no-preprocess — skip image enhancement
• --model <model> — Anthropic model (default: claude-sonnet-4-20250514)

Library

const { scanInvoice, formatOutput } = require('{SKILL_DIR}/scripts');
const invoice = await scanInvoice(filePath, { apiKey });

// Format
const json = formatOutput(invoice, 'json');
const csv  = formatOutput(invoice, 'csv');
const xlsx = formatOutput(invoice, 'excel'); // Buffer — write to .xlsx

Agent-Native Mode (No API Key)

If you have vision capabilities, skip the CLI entirely:

1. Read the image yourself using the image tool
2. Extract fields following the schema in references/canonical-schema.md
3. Run validation: require('{SKILL_DIR}/scripts/validation/arithmetic').validateArithmetic(invoice)
4. Format output: formatOutput(invoice, 'excel')

This avoids an API call — you ARE the vision model.

Conversational Use

When user sends an invoice:

1. Save the image, run scanInvoice(path, { apiKey })
2. Summarise:

📄 Invoice #12345 | 2026-02-15
   Supplier: Acme Ltd → Buyer: TechCorp UK
   Net: £1,125.00 | VAT: £225.00 | Gross: £1,350.00
   Items: 3 | Arithmetic: ✅ | Quality: good (9/9)

3. If rejected: "Classified as {type}, not an invoice."
4. Offer: "Want JSON, CSV, or Excel?"

Output Formats

JSON — full canonical schema, all fields, nested objects.

CSV — Zoho-style flat: one row per line item, header fields repeated per row. 41 columns.

Excel — Sheet 1 "Invoice": same flat table as CSV. Sheet 2 "Validation": arithmetic, errors, warnings, field confidence, stamps, handwriting annotations.

What Gets Extracted

Header: invoice#, dates, supplier/buyer (name, address, VAT#), currency, payment terms, bank details (IBAN/BIC/account/sort).

Line items: description, qty, unit, price, total, VAT %, SKU, discount.

References: PO, contract, GRN, inspection, timesheet, project, proforma.

Additional: handwritten notes, stamps/seals, document language.

Validation

• 9 document types: invoice, credit-note, receipt, purchase-order, delivery-note, confirmation, statement, other-financial, not-financial
• Arithmetic: line sums, VAT calcs, gross = net + VAT
• 15 business rules: missing fields, suspicious numbers, receipt signals
• Quality score: 0.0–1.0 (good ≥0.7, partial ≥0.3, poor <0.3)

Accept Modes

• strict — invoices + credit notes only
• relaxed (default) — also receipts, proformas
• any — extracts all, just classifies

Multi-Language & Locale

Auto-detects language. Handles: US/UK (1,234.56), European (1.234,56), French (1 234,56), Indian (1,23,456.78). Tested: English, Chinese, Russian, Filipino, German.

References

• Full schema: Read references/canonical-schema.md
• Validation rules: Read references/validation-rules.md