AI reads a logo sheet and fills linked Airtable tables

AI reads a logo sheet and fills linked Airtable tables

An n8n workflow with an upload form. You send a picture that shows the logos of several tools, such as a chart that
compares AI tools. An AI model reads the picture and lists each tool with its attributes and similar tools. The workflow
then saves them in two linked Airtable tables.

What is inside

• workflow.json: the workflow as n8n exports it, 33 nodes and 11 sticky notes. Two of the nodes are switched off, and
  two No Operation nodes only mark the table layout. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The model node ran gpt-4o. OpenAI's deprecations page, read on 12 November 2026, shuts the gpt-4o snapshots down on
  14 January 2027 and names gpt-6-luna as the replacement. This agent has an output parser connected. n8n's Tools
  Agent hands that parser to the model as a tool. gpt-6-luna on Chat Completions calls tools only with reasoning effort
  none, which n8n's OpenAI Chat Model never sends. The node now runs gpt-5.6-terra, which takes images and calls
  tools. It is renamed OpenAI Chat Model.
• Map Agent Input read a field called Prompt, but the form keyed its optional field by a long label, so the prompt
  never arrived. The field is now labelled Additional prompt, and the node reads it. When the field is empty, it sends
  a short instruction to list every tool or logo in the image.
• The Airtable base and table pickers carried the original author's saved ids and names. They are blank now.
• Two sticky notes held the original author's logo, a promotional link and a picture hosted on their site. Those lines are
  gone. The second note now describes the example sheet in words.
• A helper node, Table: Tools, pointed at itself. That link is removed.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An Airtable base with two tables, set up as shown below.
• An Airtable personal access token with the scopes data.records:read, data.records:write and schema.bases:read,
  and access to that base.
• An OpenAI API key.
• A picture with several logos on it.

The Airtable layout

• Tools: Name (single line text), Hash (single line text), Attributes (link to the Attributes table) and
  Similar (link to the Tools table). You may add Description (long text), Website (URL) and Category (multiple
  select), but the workflow does not fill them.
• Attributes: Name (single line text) and Tools (link to the Tools table).

Set it up

1. In n8n, create a workflow, open its menu, choose Import from File and pick workflow.json.
2. In Airtable, create the base and the two tables above.
3. Add an Airtable credential with your token to Check if Attribute exists, Create if not Exist, It Should exists
   and Save all this juicy data.
4. In each of those four nodes, pick your base. Pick the Attributes table in Check if Attribute exists and the
   Tools table in the other three. If n8n clears a node's field mapping when you pick the table, map the fields
   again as listed below.
5. Add an OpenAI credential to OpenAI Chat Model.
6. Save the workflow and switch it on. Open the form at /form/logo-sheet-feeder on your n8n, upload a picture and
   send it.

What each Airtable node writes

• Check if Attribute exists: Name is the attribute text, in the Attributes table. The row is matched on Name.
• Create if not Exist: Name is the tool's name and Hash is its hash, in the Tools table. The row is matched on
  Hash.
• It Should exists: the same two fields for a similar tool, in the Tools table, matched on Hash.
• Save all this juicy data: Name, Hash, Attributes and Similar in the Tools table, matched on Hash. The
  last two hold lists of record ids.

How a run goes

1. On form submission serves the form logo-sheet-feeder, with a required picture and an optional prompt.
2. Map Agent Input puts the prompt into chatInput.
3. Retrieve and Parser Agent runs on OpenAI Chat Model. It receives the picture, which the agent passes on to the
   model by default. The model names every tool it sees, with attributes and look-alikes. Structured Output Parser holds the
   shape of that list, JSON it unwraps it, and Split Out Tools makes one item per tool.
4. The attributes come first. Loop Over Attributes takes one tool at a time. Split Out each Attribute String makes
   one item per attribute, and Check if Attribute exists creates or finds each one in Airtable. Merge and
   Map Attribute ID pair every attribute with its record id, and All Attributes collects the pairs.
5. Wait for Attribute Creation releases the tools once every attribute exists. `Change each Attribute to the
   corresponding RecID` swaps each tool's attribute texts for those record ids.
6. The tools come next. Generate Unique Hash for Name makes an MD5 hash of the lowercased, trimmed tool name. Then
   Create if not Exist creates or finds the tool by that hash. After that, Only what we need keeps its hash, id and
   existing links, and Merge Old Data + RecID joins them on the hash. Last, Determine Attributes we should save adds
   the new attribute ids to the old ones.
7. The similar tools follow the same pattern. Split Out similar and Generate Unique Hash for Similar prepare them,
   and It Should exists creates any that are missing. All Similar collects the pairs, and Merge2 waits for them.
   Change each Smiliar to the corresponding RecID swaps the texts for ids, and Determine Similar we should save adds
   them to the old links.
8. Save all this juicy data writes each tool with its name, hash, attributes and similar tools.

The two Table: nodes and the two switched-off nodes, make it a readable list and Get Schema, are documentation
helpers. They take no part in a run.

Change the look

• What the agent extracts: the system message in Retrieve and Parser Agent and the example in Structured Output Parser.
• The model: OpenAI Chat Model.
• The form's fields and address: On form submission.
• More Airtable fields: add them in Save all this juicy data and ask for them in the agent's instructions.

Good to know

• The facts below come from n8n source, OpenAI pages and Airtable pages read on 12 November 2026. Nothing was run: no
  form, no model call and no Airtable base.
• Reading a picture is best effort. The model can miss a logo. Upload the same picture twice, or add a second agent that
  checks the list, if you need all of them.
• A tool's identity is the hash of its lowercased, trimmed name. "Notion" and "Notion AI" become two tools, and a
  spelling variant splits one tool into two.
• Every attribute, tool and similar tool is its own Airtable request. Airtable's pages say it may throttle upsert
  requests differently from its standard limit, so a large sheet can hit a limit.
• The form has no authentication. Anyone who has its link can upload a picture, and each upload spends OpenAI and Airtable
  calls.
• gpt-5.6-terra costs USD 2 per million tokens in and USD 12 per million out on OpenAI's page. A picture counts as
  input, and reasoning tokens count as output.
• Your data: the picture and the prompt go to OpenAI, and the names and links go to Airtable.
• Credentials: the JSON holds none.