Bank statement PDF turned into Markdown by Gemini vision

Bank statement PDF turned into Markdown by Gemini vision

An n8n workflow that reads a bank statement by looking at it. It downloads a PDF from Google Drive and has a Stirling PDF
server turn each page into a picture. Gemini writes every page out as Markdown with its tables, and a second Gemini step
lists each deposit with its date, description and amount.

What is inside

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

What we changed from the original copy

• Both Gemini steps asked for a 1.5 model that Google's model list no longer shows. They now use models/gemini-3.8-flash,
  which accepts images.
• The Stirling PDF address, the Google Drive file id and a Drive link in the notes were private. They are now the
  placeholders <YOUR_STIRLING_PDF_URL> and <YOUR_BANK_STATEMENT_FILE_ID>.
• The original sorted the page pictures by file name as text, so page 10 came before page 2. Images To List now reads the
  page number from each file name and Sort Pages sorts by that number. If a name has no number, the page keeps its place
  in the list.
• The instruction in Transcribe to Markdown had a typo, and the deposit instruction is reworded.
• Several nodes were renamed to say what they do. The wiring is unchanged. The two sticky notes are rewritten, and one
  now warns that a statement is private.

Before you start

• A Google Drive account whose Drive you can connect with OAuth2, and the PDF statement in it.
• A Stirling PDF server that you run yourself. A public one would see the whole statement.
• A Google Gemini API key.
• An n8n that has GraphicsMagick for the Edit Image step. n8n's Docker image includes it.

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 Get Bank Statement. Replace <YOUR_BANK_STATEMENT_FILE_ID> with the id
   of your PDF, the part of its share link after /d/.
3. Start your own Stirling PDF server. In Split PDF Into Page Images, replace <YOUR_STIRLING_PDF_URL> with its address,
   with no slash at the end.
4. Add a Google Gemini credential and select it in Gemini Model for Pages and Gemini Model for Deposits.
5. Run it by hand on a statement of one or two pages and read the output before you try a long one.

How a run goes

1. Get Bank Statement downloads the PDF.
2. Split PDF Into Page Images posts it to Stirling PDF and asks for one JPG for each page at 300 dpi. The reply is a zip
   file, and Extract Zip File opens it.
3. Images To List makes one item for each picture and gives it a page number. Sort Pages puts the items in page order.
4. Resize Images For AI shrinks each picture to 75%.
5. Transcribe to Markdown sends each picture to Gemini and gets the page back as Markdown, with tables as Markdown tables.
6. Combine All Pages joins the page texts into one list called pages.
7. Extract All Deposit Table Rows sends all pages together to a second Gemini model and returns a list of deposits with
   date, description and amount.

Change the result

• The statement: the file id in Get Bank Statement.
• The sharpness of the pictures: the dpi field in Split PDF Into Page Images.
• How small the pictures get: the percentage in Resize Images For AI.
• The Markdown rules: the message in Transcribe to Markdown.
• What is listed at the end: the instruction and the fields in Extract All Deposit Table Rows.
• The model: the two Gemini nodes.

Good to know

• Every page goes to Google's Gemini service as a picture. Check your bank's terms and your own rules first.
• Each page is one Gemini call, plus one for the deposit list. A long statement costs more and takes longer.
• A model can misread a digit or merge two rows. Check every amount against the PDF.
• All pages reach the last step as one text. Mindket did not check Gemini's limit for the size of that text.
• The deposit model has the safety category for dangerous content set to block nothing. This was in the original, and it
  means Gemini does not withhold an answer for that category.
• The last step uses a schema that is a list. Mindket did not run it, and if n8n objects, wrap the list in an object with one
  field.
• Mindket did not run this workflow against Google Drive, Stirling PDF or Gemini. The page-numbering code was tested only on
  sample file names on a local computer. The Stirling PDF fields and the Gemini model name follow the documentation and the
  sources that were read, not a live call.