Invoice Field Extractor

A skill the InTouch AI assistant loads to turn raw invoice text into a clean, structured JSON record.

What it does

Reads the plaintext of a single invoice and extracts the fields you actually need for AP/bookkeeping automation:

• Vendor — the issuing company (name, and address/email/phone when present), correctly distinguished from the bill-to party.
• Invoice number — pulled from whatever label the vendor uses (Invoice #, No., Reference, Bill #).
• Dates — invoice/issue date and due date, normalized to ISO YYYY-MM-DD.
• Currency — detected from symbol or ISO code; null when absent (never guessed).
• Line items — per-row description, quantity, unitPrice, amount, with summary rows excluded.
• Totals — subtotal, tax, and the final total due, as plain decimals.
• Warnings — a list of ambiguous or unresolved fields, so nothing is silently fabricated.

What to feed it

The full text of one invoice — from OCR output, a PDF text dump, an email body, or pasted text. Field labels, date formats, and amount formatting can vary; the skill normalizes them. Run one invoice per invocation.

What it returns

A single JSON object (and nothing else) with this shape:

{
  "vendor": { "name": "...", "address": "...", "email": "...", "phone": "..." },
  "invoiceNumber": "...",
  "invoiceDate": "YYYY-MM-DD",
  "dueDate": "YYYY-MM-DD",
  "currency": "USD",
  "lineItems": [
    { "description": "...", "quantity": 1, "unitPrice": 100.00, "amount": 100.00 }
  ],
  "subtotal": 100.00,
  "tax": 8.00,
  "total": 108.00,
  "warnings": []
}

Missing scalar fields come back as null; numbers are plain decimals with no currency symbols or thousands separators. The output is ready to pipe into a downstream InTouch task (SQL insert, sheet append, accounting webhook, etc.).