Extract Invoice Data with LlamaParse & GPT-4o

Accept the invoice PDF from either email attachments (recommended for vendors) or a manual upload folder.

Actionable Prompt / Code:

System: You are an operations assistant. Your job is to capture invoice intake metadata.
Return ONLY JSON with: {"source":"email|upload","received_at":"ISO-8601","sender":"string","filename":"string","notes":"string"}.
If a field is unknown, return null.

Pro Tip: Save the raw file immediately so you always have an audit trail, even if extraction fails later.

Send the saved PDF to LlamaParse and request a layout-faithful output (markdown or text) so line items and totals stay readable.

Actionable Prompt / Code:

import os
from llama_parse import LlamaParse

parser = LlamaParse(
    api_key=os.environ.get("LLAMA_CLOUD_API_KEY"),
    result_type="markdown",
    verbose=False,
)

# file_path = "/path/to/invoice.pdf"
# documents = parser.load_data(file_path)
# parsed_markdown = "

".join([d.text for d in documents])

Pro Tip: If the invoice is multi-page, keep page separators so you can trace fields back to the source page during review.

Give the parsed text to GPT-4o and require strict schema output so totals, taxes, and line items are normalized.

Actionable Prompt / Code:

System: You are a finance data extraction engine.
You must return ONLY valid JSON that matches the schema exactly. No markdown, no commentary.

Normalization rules:
- Dates must be ISO-8601 (YYYY-MM-DD).
- Amounts must be numbers (no currency symbols).
- Currency must be an ISO 4217 code when possible (e.g., USD, JPY, EUR).
- Line items must include quantity and unit_price when present; otherwise null.

JSON Schema (strict):
{
  "invoice_number": "string|null",
  "invoice_date": "string|null",
  "vendor_name": "string|null",
  "vendor_tax_id": "string|null",
  "bill_to": "string|null",
  "currency": "string|null",
  "subtotal": "number|null",
  "tax": "number|null",
  "total": "number|null",
  "due_date": "string|null",
  "payment_terms": "string|null",
  "line_items": [
    {"description":"string|null","quantity":"number|null","unit_price":"number|null","amount":"number|null"}
  ]
}

Pro Tip: Reject outputs when subtotal + tax differs from total by more than 0.5% to catch OCR/parse drift early.

Append the JSON into Google Sheets as a single row (one invoice per row), then store the original PDF in Google Drive and keep the file URL in the sheet.

Actionable Prompt / Code:

// Pseudo-code (works with any Sheets/Drive SDK)
const row = [
  data.invoice_number,
  data.invoice_date,
  data.vendor_name,
  data.currency,
  data.subtotal,
  data.tax,
  data.total,
  data.due_date,
  drive_file_url
];
// sheets.appendRow("Invoices", row)

Pro Tip: Use a deterministic idempotency key like vendor_name + invoice_number + total to avoid duplicates.

Send a concise message via Telegram with vendor, total, due date, and the Drive link, so a human can approve in under ~10 seconds.

Actionable Prompt / Code:

Message template:
New invoice ready for review
- Vendor: {vendor_name}
- Total: {currency} {total}
- Due: {due_date}
- Link: {drive_file_url}
Reply with: APPROVE or REJECT:{reason}

Pro Tip: Track approvals in the sheet with a simple status column (Pending/Approved/Rejected) to make reconciliation deterministic.