Baserow columns filled from PDFs by field descriptions

Baserow columns filled from PDFs by field descriptions

An n8n workflow that treats each Baserow field's description as a prompt. When someone attaches a PDF in a row's File
field, a chat model reads it and fills every described field that is still empty. When a field is added or changed, the
same prompt runs again for the rows that have a PDF.

What is inside

• workflow.json: the workflow as n8n exports it, forty-four nodes: thirty that do the work and fourteen sticky notes. It
  holds no credential.
• README.md: these notes.

What we changed from the original copy

• The credential links are gone: the Header Auth credential on five HTTP Request nodes and the OpenAI key on both chat
  models. Select your own.
• Baserow Event had its author's own path and webhook ID. The ID is gone and the path is now YOUR_WEBHOOK_PATH.
• List Table API asked Baserow for six named fields only, and File was not one of them. Baserow's API schema says that
  a request with include returns just the fields it names, so the rows would have arrived with no attachment to download.
  Mindket's copy drops include.
• Get Prompt Fields passed on each field's name, order and description but not its type, so the "output format" line in
  the row branch's prompt came out empty. It now passes type too, as the field branch already gets it from the webhook.
• The large note at the top, with the author's links and a demo video, is now a title. The note that held a logo is gone,
  and the screenshot link in the webhook note is removed. The text of that note stays.
• Nothing else. The Code nodes needed no other repair.

Before you start

• An n8n instance, on n8n Cloud or your own server. Baserow must be able to reach its webhook address.
• A Baserow table with a file field named File. The name is written into the filter in List Table API, into the
  File[0].url expressions and into Filter Valid Rows, so change it in all of them or none.
• A Baserow database token that can read and update that table.
• An OpenAI API key.
• PDFs that have text in them. A scanned page with no text layer gives the model nothing to read.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Header Auth credential. Set its name to Authorization and its value to the word Token, a space and your
   Baserow database token. Select it in the five nodes that call Baserow: Table Fields API, List Table API,
   Get Row, Update Row and Update Row1.
3. Add an OpenAI credential and select it in OpenAI Chat Model and OpenAI Chat Model1. Neither names a model, so
   choose one in each.
4. In Baserow Event, replace YOUR_WEBHOOK_PATH with a long random string of your own. Save, then press Publish.
5. In Baserow, open the table's webhook settings from the three-dots button in its toolbar. Create a POST webhook to the
   production address of Baserow Event. Tick the option to use field names instead of IDs. Choose individual events:
   "row updated", "field created" and "field updated". For "row updated", pick the File field as the one to watch.
6. In Baserow, write a prompt in the description of each field you want filled, such as "The company that sent the
   document". Attach a PDF to a row. The empty described cells fill within moments.

A self-hosted Baserow needs one more change: replace https://api.baserow.io with your own address in the five nodes above.

How a run goes

1. Baserow Event receives the event. Table Fields API lists the table's fields, Get Prompt Fields keeps those with a
   description, and Get Event Body puts the event's own data back in front. Event Type sends rows.updated one way and
   field.created and field.updated the other.
2. A row change: Rows to List splits the event into rows, Filter Valid Rows keeps those with a file, and Get Row
   reads each one fresh. Loop Over Items1 takes one row. The PDF is downloaded and read as text. Fields to Update
   lists the described fields whose cell is empty. For each one, Generate Field Value1 sends the PDF's text, the
   field's description and its type to the chat model. Get Result1 names the answer after the field and Update Row1
   patches the row.
3. A field change: Filter Valid Fields keeps it if the field has a description. List Table API lists the rows that
   have a file, Get Valid Rows flattens its pages, and Loop Over Items takes one row. The PDF is read,
   Generate Field Value answers for that one field, and Update Row patches the row.

Change the look

• The prompts: the Baserow field descriptions, and the system messages of Generate Field Value and
  Generate Field Value1, which keep answers short and ask for n/a when nothing is found.
• The models: both chat model nodes.
• The input column: the name File, in the places listed above.
• How many rows a field change reaches: Max Pages in the pagination of List Table API, and its size.
• Which edits start a run: the events and the watched field in Baserow's webhook settings.

Good to know

• A row change fills only empty cells. A field change fills that column for the rows it fetches, and replaces what was
  there.
• List Table API asks for 20 rows a page and stops after 3 requests, so a field change reaches at most 60 rows. Raise
  the page limit to reach more. The note beside it warns that a table of thousands of rows would take a long time.
• Each field on each row is one chat call that carries the whole PDF's text. Many fields on long PDFs cost more.
• Only the first attachment, File[0], is read, and Extract from File is set to PDF.
• The answer n/a is text. Baserow can refuse text in a number or date field. Describe text fields first.
• Baserow's webhook watches only the File field, so the workflow's own writes to the other columns do not start
  another run.
• Baserow's API schema, read on 12 November 2026, lets a database token call the fields and rows endpoints this workflow
  uses, written as Token and the token.
• Neither chat model names a model. OpenAI Chat Model 1.1 uses gpt-5-mini in n8n's current source, read on 12 November
  2026. OpenAI's deprecations page gives 11 December 2026 for gpt-5-mini and names gpt-5.6-terra as its replacement.
• Keep the webhook path long and private. Anyone who learns the address can call it.
• The PDF's text goes to OpenAI. Your rows travel between Baserow and n8n.