AI Prompt Node
What it does
The AI Prompt node sends a prompt, filled in with each row's values, to an AI model and adds the model's answer to the row. It classifies rows: the answer is a category and, when the model gives one, a confidence. Rows leave on out, or on Review when the answer needs separate handling.
The node is in the AI group of the component browser.
When to use
- Rows hold free text, such as support tickets, product descriptions or emails, and each needs a label.
- No rule decides the label reliably.
- You want a separate path for rows the model is unsure about.
Each row is one model call and counts toward the account's AI usage. Filter rows first so the model only sees the rows a rule cannot decide. Usage is shown under Manage Account > Billing & Usage.
Configure
Select the node and press Configure in the side panel to open Configure AI Prompt. Save keeps the dialog open, Done saves and closes it, and Cancel discards the changes since the last save. A message confirms each save.
- Prompt mode:
Classify. The other modes are listed and cannot be selected. - Prompt template: the prompt sent for each row.
{{column}}inserts the row's value, for exampleClassify the sentiment of: {{message}}.
Model and review
- Model override: leave empty to use the default model.
- Review category: a result that matches this category, ignoring case and surrounding spaces, goes to
Review. When the model call fails, the row also goes toReviewwith this category. - Confidence threshold: a number above 0 and up to 1. A result below it goes to
Review. A result with no confidence is not affected.
Ports
| Port | Direction | Kind | Notes |
|---|---|---|---|
in | in | data | Rows to classify. |
out | out | data | Rows with the model's answer. |
Review, id fallback | out | data | Rows to handle separately. The ai_fallback_reason column says why. |
error | out | control | Rows the node could not classify, with the reason. |
Output columns
Columns added to each row
| Column | Meaning |
|---|---|
category | The model's answer. |
confidence | The model's confidence from 0 to 1, or empty when it gave none. |
ai_provider, ai_model | Which model answered. |
ai_input_tokens, ai_output_tokens | Token counts for the call. |
ai_fallback_reason | On Review only: category_match, low_confidence, provider_error or entitlement_denied. |
Where each row goes
| Result | Port |
|---|---|
| A category that is not the review category, at or above the threshold | out |
| The review category | Review, reason category_match |
| Below the confidence threshold | Review, reason low_confidence |
| The model call failed and a review category is set | Review, reason provider_error |
| The account has used its AI credits for the month | Review, reason entitlement_denied |
| The model call failed and no review category is set | error |
| The prompt is empty for the row | error |
| The model's answer could not be read | error. The call still counts toward usage. |
Connect Review. Rows on an unconnected port go no further.
Examples
Route support tickets.
Prompt Categorise this ticket as billing, technical or sales: {{body}}, review category unknown, confidence threshold 0.7. A Router sends each category to its own queue. Rows the model is unsure about go to a Notification for the team to review.
Review only borderline expenses. Source > AI Prompt (classify as auto or manual) > Router > manual rows to a Human Approval node.
Troubleshooting
| Symptom | Cause | What to do |
|---|---|---|
Rows reach error with prompt_empty | The template is empty, or every column it uses is empty on that row. | Check the template and the column names. |
Rows reach Review with entitlement_denied | The account has used its AI credits for the month. | Check AI credits under Manage Account > Billing & Usage. Credits reset at the start of the next billing period. |
Many rows reach Review with low_confidence | The threshold is high, or the categories overlap. | Lower the threshold, or name the categories more clearly in the prompt. |
Rows reach error with invalid_json | The model's answer was not in the expected form. | Ask for one of the listed categories in the prompt. |
| Rows stop at the node | Review is not connected. | Connect it. |