---
title: Guardrails
description: Validate text against policies using a model, and route to pass or fail
---

Ask a chat model whether a piece of text follows your rules, then send the data down a **Pass** or a **Fail** path.

## What it does

New Guardrails steps (version 2) run **quick checks** first, with no AI and no network, then an optional **AI policy**:

* **Personal data**: emails, phone numbers, IBANs, card numbers (checked with the Luhn rule), IPv4 addresses, and postal addresses (best effort).
* **Secrets**: API keys and tokens, *balanced* or *strict*.
* **Links**: only allowed sites and schemes, and no passwords inside URLs.
* **Patterns** and **keywords** of your own.
* **AI policy** (optional): the chat model in the **Model** slot judges the text against your policy and answers with `{ flagged, confidence, reason }`. The step fails the text at or above the **threshold**. The text is sent fenced, with the instruction to judge it and never follow it, so a text can't talk the model into passing it. An answer that can't be read **fails closed**.

The **Operation** decides what happens:

* **Check**: the data goes out of **Pass**, or out of **Fail** at the first finding, with what was found.
* **Clean up the text**: findings are replaced with `[EMAIL]`, `[SECRET]`, `[URL]` or your label, and the cleaned text goes out of one **Output**. The AI policy, if on, only ever sees the cleaned text.

The model is only needed for the AI policy.

### Steps saved before AWFlow 0.9

They send the **Text** and your **Policy** to the model, which answers passed or not, with a reason: **Pass** or **Fail**, failing closed. **Update** in the step's header turns your policy into the AI policy check (failing above 0.7 confidence) and leaves the quick checks off, so results stay the same until you turn some on.

## When to use it

* Check an AI-written reply before you post or send it (no personal data, no promises, stays on topic).
* Screen text a user typed or selected before it reaches another AI step.
* Stop a workflow when scraped content is off topic or unsafe.

To sort text into more than two groups, use [Text Classifier](/nodes/builtin/ai/aiagents/textclassifiernode/). For a plain rule that needs no AI (a word is present, a number is above a limit), use [If](/nodes/builtin/flow/if/).

## Inputs and settings

Fill **Text** with an expression, for example `{{ $input.response }}` to check the reply of an earlier AI node.

Write **Policy** as clear, checkable rules, for example: `No email addresses or phone numbers. No profanity. The text must be about our product's shipping times.` Vague policies give inconsistent verdicts.

## Outputs

What comes out depends on whether something is connected to the node's input:

* **With an incoming connection**, the incoming items are passed on unchanged to **Pass** or **Fail**. The verdict's reason is not added to them.
* **With nothing connected** to the input, the node sends one item `{ text, passed, reason }`, so you can read why the text failed.

## Dependencies and credentials

Connect a chat model such as [Chat OpenAI](/nodes/builtin/ai/aidependencies/llm/openai/) or [Chat Anthropic](/nodes/builtin/ai/aidependencies/llm/anthropic/). Guardrails has no credential of its own: the model node holds the API key.

## Example workflow

1. [Get Selected Text](/nodes/extension/getselectedtext/) reads the customer message you highlighted.
2. [Basic LLM Chain](/nodes/builtin/ai/aiagents/basicllmchainnode/) drafts a reply with **Prompt** `Write a short, friendly reply to: {{ $input.text }}`.
3. Guardrails with **Text** `{{ $input.response }}` and **Policy** `No refunds or discounts may be promised. No personal data.`, with an [Chat OpenAI](/nodes/builtin/ai/aidependencies/llm/openai/) model on **Model**.
4. From **Pass**, [Clipboard Write](/nodes/extension/clipboardwrite/) copies the reply. From **Fail**, [Show Notification](/nodes/extension/shownotification/) tells you to write it by hand.

## Common issues

* **Everything goes out of Fail.** The model probably adds text around its verdict, which counts as a failure. Use a stronger model, or make the policy shorter and more precise. Run the node with nothing connected to its input to see the `reason`.
* **"Dependency model is required".** Connect a chat model to the **Model** slot.
* **"Model not found".** The connected model node is not set up: open it and choose a model and its credential.
* **Verdicts change from run to run.** The policy is open to interpretation. List concrete rules instead of general goals.

## Related nodes

* [Text Classifier](/nodes/builtin/ai/aiagents/textclassifiernode/): sort text into several categories.
* [Sentiment Analysis](/nodes/builtin/ai/aiagents/sentimentanalysisnode/): route by tone.
* [IF](/nodes/builtin/flow/if/): branch on a simple rule, no AI needed.
* [Basic LLM Chain](/nodes/builtin/ai/aiagents/basicllmchainnode/): generate the text you want to check.