Watch Me Build A PDF Invoice Parsing Flo

Watch Me Build A PDF Invoice Parsing Flow In Minutes

Before you run this

By default this skill uses local CSV files for its data (input and output) — nothing to connect, works offline, your data stays on your machine. The CSVs live next to the skill in ./data/ (input: data/input.csv, output: data/output.csv), or point it at any path you like.

If you'd rather read/write a Google Sheet instead, just say so and I'll switch you over. It's a one-time setup: you'll paste a Google service-account JSON (or authorize once), share your Sheet with that account's email, and give me the Sheet URL. After that it behaves exactly the same, just backed by your Sheet. Say "use Google Sheets" to start that, or "keep CSVs" (default) to just go.

What this does

Reads a list of PDF invoice URLs from a CSV, extracts the raw text from each PDF using PDF.co, then passes that text to GPT-4o to pull out three structured fields: total amount, line items, and sender address. Writes the enriched rows back out to data/output.csv.

This is the Python translation of the "Parse Invoice -> Add to Sheet" Make scenario: same logic, same prompt, no OAuth needed by default.

When to trigger this

• You have a batch of invoice PDFs (as URLs) and want to extract structured data without manually opening each one
• You're building an accounts-payable or expense-tracking automation and need clean totals + line items
• You want to see how PDF text extraction + LLM parsing work together in a real pipeline

Required env vars

PDFCO_API_KEY=your_pdf_co_key
OPENAI_API_KEY=your_openai_key

Set these in your .env or export them before running.

Input CSV format

data/input.csv — one row per invoice, one column:

Rows with an empty invoice_url are skipped.

Step-by-step procedure

1. Add your invoice URLs to data/input.csv (one URL per row under the invoice_url header).
2. Set PDFCO_API_KEY and OPENAI_API_KEY in your environment.
3. Run the pipeline:

   python scripts/parse_invoices.py
   

4. Results land in data/output.csv with columns: invoice_url, total, line_items, from_address.
5. If a PDF fails to convert or GPT returns malformed JSON, that row gets error logged and the script continues — you won't lose the whole batch.

Node map (source Make scenario -> this skill)

Conditional page logic (source node 6): if the URL contains "canadapost", only page 0 is extracted. This is preserved as-is in parse_invoices.py.