Vision

Look at an image and answer questions about it.

What it does

Looks at an image and answers a question about it. This is how a photograph becomes data — a receipt, a meter, a delivery note, a damaged parcel.

What you need

API key from OpenAI. platform.openai.com → API keys. This is for the vision and image tools only — the key your agents think with is set separately, under LLM Connections.

Connecting it

Credentials belong to one project, not to your workspace and not to you. A system built for one client never runs on another client's account, and a package handed to a business runs on theirs.

1. Open the project and stay on the Visual Editor tab.
2. Open the Tools & Integrations rail on the right.
3. Find this tool, open it, and paste the credential.

To change it later, open the same panel — you can leave the key blank to keep the stored one and change only the other field.

Using it in a step

An Action step runs one tool with the arguments you type, exactly as written, with no model involved — so it costs nothing per run and behaves the same every time. Add one from the canvas toolbar, choose Vision, and fill in:

Values can carry results forward: {{previous}} is the step before, and {{row:column}} is a column from the row this run is about when the trigger fans out.

Press Try it to run the step on its own before committing to a whole sequence.

Using it with an agent

Tick Vision on any agent that should be able to call it, and the agent decides when to. Use an agent when the judgement matters and an Action step when the certainty does.

A worked example

A distributor's drivers photograph signed delivery notes. The agent reads the customer name and the date off each one and files it against the order.

When it goes wrong

It misreads handwriting. Ask a narrow question — 'what is the total?' beats 'read this receipt' — and expect to check anything financial.

See also

• Action steps — the step kind this is used in
• Sequences — how steps fit together
• Triggers — running it on a schedule