Drive invoices parsed by LlamaParse and saved in Airtable

Drive invoices parsed by LlamaParse and saved in Airtable

An n8n workflow that files invoices. It watches a Google Drive folder, sends each new file to LlamaParse, and receives the parsed text on a webhook. An OpenAI model then returns the line items as JSON, and Airtable gets one record for the invoice and one for each line item.

What is inside

• workflow.json: the workflow as n8n exports it, 9 nodes beside 9 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The original file had stray text after its JSON. Mindket’s copy drops it.
• The upload went to https://api.cloud.llamaindex.ai/api/parsing/upload. LlamaParse’s documentation, read on 12 November 2026, shows /api/v1/parsing/upload, and Mindket’s copy uses that.
• The parsing_instruction was sent as a header. The documentation lists it as a form field, so it is now in the form.
• The OpenAI step read body.json[0].items from the webhook. The documented webhook body has txt, md, json and images, and the json sample has no items. Mindket’s copy sends body.md, the text as Markdown, and the prompt now says what to return.
• The model gpt-4o-mini is shut down on 14 January 2027, says OpenAI’s deprecations page, with gpt-6-luna as its replacement. Mindket’s copy uses gpt-6-luna. It has no tools here, so the note about tools on that model does not apply.
• Amounts such as $1,250.00 lost their thousands part, because parseFloat stops at the comma. The two price expressions now remove the dollar sign, commas and spaces first.
• The webhook path, the webhook address, the Drive folder, the Airtable base and tables, and the empty key in the header are placeholders. The original held its author’s own ids.
• The OpenAI node had a spaced hyphen in its name. It is now OpenAI Extract Line Items, and the Code node that reads it follows.
• The sticky notes are reworded. The banner note described a different workflow, a meeting agent, and is replaced.
• No credential is inside.

Before you start

• An n8n instance with a public HTTPS address and a domain name, on n8n Cloud or your own server.
• A Google account and a Drive folder for the invoices.
• A LlamaCloud API key.
• An OpenAI API key.
• An Airtable base with two tables: Invoices, with a Name column and a Line Items link, and Line Items, with Description, Qty, Unit price, Amount and an Invoices link. Add an Airtable personal access token.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Google Drive credential and select it in Google Drive Trigger and Google Drive. Choose your invoice folder in the trigger.
3. In Webhook, replace <YOUR_RANDOM_WEBHOOK_PATH> with a long random text. Use the same text in the address in Upload File, with your n8n host in place of <YOUR_N8N_HOST>.
4. In Upload File, replace <YOUR_LLAMACLOUD_API_KEY> in the Authorization header with your key.
5. Add an OpenAI credential and select it in OpenAI Extract Line Items.
6. Add an Airtable credential, then choose your base and your two tables in Create Invoice and Create Line Item.
7. Publish the workflow and drop an invoice PDF into the folder.

How a run goes

1. Google Drive Trigger checks the folder every minute and starts a run for each new file. Google Drive downloads it.
2. Upload File posts the file to LlamaParse with the webhook address and an instruction to find the line items.
3. When LlamaParse is done, it posts the result to Webhook. Set Fields adds the prompt and the JSON schema.
4. OpenAI Extract Line Items asks the model for the lines, in the shape of the schema.
5. Create Invoice adds an empty record to the Invoices table.
6. Process Line Items turns the model’s JSON into one item for each line. Create Line Item adds each line to the Line Items table, linked to the invoice.

Change the look

• The folder that is watched: Google Drive Trigger.
• What LlamaParse is asked to do: the parsing_instruction field in Upload File.
• Reading scanned pages: remove the disable_ocr field in Upload File.
• What the model returns: the prompt and the schema in Set Fields.
• The columns that are filled: the two Airtable nodes.

Good to know

• LlamaParse’s documentation, read on 12 November 2026, calls its first API version deprecated but still available, and says new features come to the second version only. This copy stays on the first one, whose webhook is documented to send the result. The second version changes the webhook, so test it before you move.
• LlamaParse takes only an HTTPS webhook address with a domain name, under 200 characters. A local n8n needs a tunnel.
• Anyone who learns the webhook path can post text to it, and that text goes to OpenAI and into Airtable. Use a long random path.
• disable_ocr is true, so text inside images is not read. A scanned invoice may come back empty.
• Create Invoice fills no column. Name the record in Airtable, or map a field.
• A European price such as 1.234,50 is not read correctly by the two price expressions.
• The invoice text goes to LlamaParse, OpenAI and Airtable, so check that you may send it.
• Mindket did not run this workflow. It read LlamaParse’s and OpenAI’s documentation and n8n’s source instead.
• The JSON holds no credential.