Survey photos named by an AI agent and saved to Airtable

Survey photos named by an AI agent and saved to Airtable

An n8n workflow that fills in a survey table from its photos. It reads the Airtable rows that have a photo and are not done yet. A vision model describes each photo, and an AI agent searches the web with a reverse image search and a page scraper to name the object, its model, material, color and condition. The answers are written back to the same row.

What is inside

• workflow.json: the workflow as n8n exports it, 29 nodes. Nineteen of them run and ten are sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• Analyse Image was an OpenAI node on version 1.3. In n8n's source, read on 12 November 2026, that version always sends gpt-4-vision-preview, which OpenAI's deprecation page lists as shut down on 6 December 2024. The node is on version 1.8 now and names gpt-6-luna, which takes image input.
• The node sets its answer length to 1500 tokens. The default of 300 is small for a model that reasons before it answers.
• OpenAI Chat Model1 used gpt-4o with a temperature of 0. OpenAI's page, read on 12 November 2026, shuts gpt-4o down on 14 January 2027 and names gpt-6-luna as the replacement. That page says Chat Completions supports function calling only with reasoning effort set to none, and n8n's chat model node offers only low, medium and high. The agent calls tools, so the node uses gpt-5.6-terra, whose page lists function calling without that limit. It has no temperature setting, because that page lists none.
• Object Identifier Agent was an OpenAI Functions agent on version 1.5. n8n's source marks that mode as deprecated and removed in n8n v3. It is a plain AI Agent on version 2.2 now, with the same prompt, tools and output parser.
• With the new agent, the answer sits one level higher. Enrich Product Rows reads {{ $json.output.color }} where it read {{ $json.output.output.color }}, and the same for the other five fields.
• Firecrawl Scrape API called /v0/scrape, which Firecrawl's server now answers with a notice to use /v2/scrape. It calls /v2/scrape now. The reply still holds the page text in data.markdown.
• The Airtable base and table IDs are placeholders, <YOUR_AIRTABLE_BASE_ID> and <YOUR_SURVEY_TABLE_ID>. The credential links are gone, and the file's name is ours.
• Nothing else. The prompts, the router, the tools and the sticky notes are as the file had them.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key.
• A SerpApi account and key, for the reverse image search.
• A Firecrawl account and key, for the page scraper.
• An Airtable account with a personal access token and a base that holds a survey table.

The survey table needs these fields, spelled exactly like this:

• Title: text. It names the object in the photo, and the agent writes its own title back here.
• Image: an attachment field with the photo.
• AI_status: a checkbox. A row with it ticked is skipped.
• Description, Model, Material, Color and Condition: text.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Open Get Applicable Rows and Enrich Product Rows. Add an Airtable credential and replace the two placeholders with your base and table. A base ID starts with app and a table ID with tbl, and both are in the Airtable address.
3. Add an OpenAI credential and select it in Analyse Image and OpenAI Chat Model1.
4. Add a SerpApi credential and select it in SERP Google Reverse Image API.
5. In Firecrawl Scrape API, create a Header Auth credential. Name the header Authorization and give it the value Bearer followed by your Firecrawl key.
6. Put a few rows with a photo and an empty AI_status in the table. Three rows make a fair first test.
7. Press Test workflow and check the rows afterward.

What happens in a run

1. When clicking "Test workflow" starts the run by hand.
2. Get Applicable Rows returns the rows that have an image and whose AI_status is not ticked.
3. For each row, Analyse Image sends the large thumbnail of the first photo and the row's Title to the vision model. It asks for a description, the model or make, the materials, the colors and the condition, and tells the model to leave an attribute blank when it cannot tell.
4. Object Identifier Agent receives the photo address and that description. It may call the reverse image search tool and the web page scraper tool as often as it needs, and it returns title, description, model, material, color and condition through the structured output parser.
5. Both tools call this same workflow. Execute Workflow Trigger takes the call, Edit Fields keeps the route and the query, and Actions Router sends it down one of three paths.
6. The reverse image path asks SerpApi for pages that hold similar images and returns each one's position, title, link and snippet. The scraper path asks Firecrawl for a page and returns the first 3000 characters of its text, or a message that the scrape failed. Any other route gets a message that the tool was not found.
7. Enrich Product Rows updates the row with the six answers and ticks AI_status.

What to change

• The photos you start from: the formula in Get Applicable Rows.
• What the agent looks for: the prompt in Analyse Image and the instructions in Object Identifier Agent. The output fields are in Structured Output Parser.
• The way the run starts: swap the manual trigger for a schedule or an Airtable trigger.
• More tools for the agent: add a route to Actions Router and a tool node that sets it.
• The model: OpenAI Chat Model1 for the agent and Analyse Image for the photo.

Good to know

• The agent's answers overwrite the row, including Title. Try it on a copy of your table first.
• A guess is still a guess. The agent is told to leave an attribute blank when it cannot find it, and a person should check the rows that matter.
• Airtable attachment links expire after a while. Run the workflow soon after it reads the rows.
• The photo address is sent to OpenAI and to SerpApi, which searches Google with it, and the pages the agent opens go to Firecrawl. Do not run it on photos that show people or places you may not share with those services.
• Cost: OpenAI charges for the photo and for the agent's calls, SerpApi for each reverse search and Firecrawl for each scrape.
• The scraper returns only the first 3000 characters of a page, so a long page can lose the detail you wanted.