Payment transactions from PDFs and images exported to CSV

Payment transactions from PDFs and images exported to CSV

An n8n workflow that watches a Google Drive folder. When a PDF or an image lands there, an AI model reads the payment transactions in it
and adds a category to each one. The result is saved as a CSV file in another Drive folder. PDFs go through OpenRouter, and images go
through Google's Gemini.

What is inside

• workflow.json: the workflow as n8n exports it, 16 nodes. 12 of them run and four are sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• Send data to A.I. named meta-llama/llama-3.1-70b-instruct:free. OpenRouter's public model list, read on 12 November 2026, no longer
  shows that name. The node now names meta-llama/llama-3.1-70b-instruct, which the list shows with a price per token.
• Google Gemini Chat Model named models/gemini-1.5-pro-latest, which Google's model and deprecation pages no longer list. It now
  names models/gemini-3.6-flash.
• The PDF text went into the request through encodeURIComponent, which turns it into percent codes. It now goes in with
  JSON.stringify, so the model reads plain text and a quote mark in a statement cannot break the request.
• The two Convert to CSV nodes made a CSV from the model's whole answer object, so the transactions ended up in one cell. They are now
  Save CSV text as file and Save CSV text as file from image. They write the answer text, which is already CSV, into a file.
• The upload name was {{$today}}. It is now the day's date followed by .csv.
• The author's two folder ids are now <YOUR_INPUT_FOLDER_ID> and <YOUR_OUTPUT_FOLDER_ID>.
• Vertex A.I. extract text is now Gemini extract text from image. The node it uses calls Google's Gemini API with a Google
  Gemini(PaLM) API key, not Vertex AI, and its note said otherwise.
• The sticky notes lost a link to the author's blog and an example service account email, which is now <YOUR_SERVICE_ACCOUNT_EMAIL>.
• Every credential link is gone. Select your own.
• Nothing else. The prompts, the routing and the polling are as they were.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google service account with the Google Drive API turned on, and its key.
• Two Google Drive folders, one for the PDFs and images and one for the CSV files, both shared with the service account.
• An OpenRouter account with an API key and some credit, for the PDFs.
• A Google Gemini API key, for the images.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Google Service Account credential and select it in Get PDF or Images, Download PDF, Download Image,
   Upload to Google Drive and Upload to Google Drive1.
3. In Get PDF or Images, replace <YOUR_INPUT_FOLDER_ID> with the id of the folder to watch. It is the last part of the folder's address
   in Drive. In both upload nodes, replace <YOUR_OUTPUT_FOLDER_ID> with the id of the CSV folder.
4. Add a Header Auth credential with the name Authorization and the value Bearer <YOUR_OPENROUTER_API_KEY>. Select it in
   Send data to A.I..
5. Add a Google Gemini(PaLM) API credential with your key and select it in Google Gemini Chat Model.
6. Click Test workflow, then drop a PDF with a few transactions into the input folder. Wait for the next poll and open the CSV folder.
   Then save and press Publish at the top of the canvas (or Shift+P).

How a run goes

1. Get PDF or Images checks the input folder every minute for a new file.
2. Route based on PDF or Image sends a PDF down its first output and any file whose type contains image/ down its second. Other
   files stop there.
3. For a PDF, Download PDF fetches the file and Extract data from PDF reads its text.
4. Send data to A.I. posts the text to OpenRouter and asks for all the transactions as CSV, with a column called category.
5. Save CSV text as file writes the answer into a file named transactions.csv, and Upload to Google Drive stores it in the output
   folder under the day's date.
6. For an image, Download Image fetches the file, and Gemini extract text from image sends it to Google Gemini Chat Model with a
   similar instruction.
7. Save CSV text as file from image and Upload to Google Drive1 do the same for the image's answer.

Change the look

• The folders: Get PDF or Images and the two upload nodes.
• The instructions: the message in Send data to A.I. for PDFs, and the two messages in Gemini extract text from image for images.
• The models: the model in the request of Send data to A.I., and Google Gemini Chat Model.
• The name of the saved file: the Name in the two upload nodes.
• How often it checks: the poll time in Get PDF or Images.

Good to know

• Not run: Mindket made no call to Google Drive, OpenRouter or Gemini. It ran the changed request expression on its own machine and
  read n8n's source for the file nodes.
• A model writes the CSV, so it can miss a line, misread a figure or add a remark before the header. Open each file and check the numbers.
• Extract data from PDF reads the text inside a PDF. A scan with no text in it gives nothing, so save those pages as images.
• The whole text goes in one request. The model's context is 131072 tokens on OpenRouter's list, and a longer statement can fail.
• The model is now a paid one, so each PDF costs a little on OpenRouter.
• The file goes to Drive as plain text with a .csv name. Two files made on one day get the same name.
• Only files that are created in the folder after the workflow starts are picked up.
• Your data: the text of a PDF goes to OpenRouter and the provider behind it, and an image goes to Google. Statements hold account
  numbers, so check both services' terms first. The CSV files stay in your Drive.
• Credentials: the JSON holds none.