Ctrlk
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Guard Nodes

Check workflow inputs and generated answers, then route execution based on the result.

Guard nodes let you check inputs and generated answers before your workflow continues. Use them to detect prompt attacks, check for unsafe content, or verify that an AI-generated answer is supported by its source material.

With Guard nodes, you can:

  • Check incoming text before sending it to an LLM.

  • Check generated answers before using them in subsequent actions.

  • Choose between Prompt guard, Content guard, and Hallucination guard.

  • Define separate paths for Not triggered and Triggered results.

Guard types

Guard type
Use it to

Prompt guard

Check incoming text for attempts to manipulate the LLM’s behaviour.

Content guard

Check incoming text or generated answers against content categories or custom rules.

Hallucination guard

Check whether an AI-generated answer is supported by its source material.

Each Guard node performs one type of check. To apply multiple checks, add multiple Guard nodes to your workflow.

Configuration guide

  1. Add a Guard node from the System section in the workflow builder.

  2. Connect it to the preceding node.

  3. Select a Guard type.

  4. Configure the content to check.

  5. Connect the Not triggered and Triggered branches to the appropriate next steps.

For text inputs, you can select an output from a connected upstream node or compose text using variables.

Prompt guard

A Prompt guard checks incoming text for attempts to manipulate the LLM’s behaviour, such as instructions to ignore its original prompt or perform an unintended action.

Place it before the LLM node that will process the text.

Text to check

Required. Provide the text you want to inspect. Typical inputs include:

  • A customer’s support ticket.

  • An incoming email.

  • A message submitted through a form.

  • Text received through a webhook.

Example: Check an incoming ticket

In a ticket-classification workflow, select the incoming ticket description as Text to check.

  • Connect Not triggered to the LLM node that classifies the ticket.

  • Connect Triggered to an alternative path, such as flagging the ticket for review.

Content guard

A Content guard checks text against content categories or custom rules. Use it to check incoming content before an LLM processes it, or to check an AI-generated answer before sending it.

Text to check

Required. Provide the text you want to inspect. For example, select an incoming message or the text output from an LLM node.

Role

Specify whose text is being checked. For example:

Role
Use for

user

Incoming messages.

assistant

Generated replies.

Leave this field empty to use the default role, user.

Content categories

Select the categories you want the guard to check. Leave the selection empty to check all categories.

Available content categories

  • Violent Crimes

  • Non-Violent Crimes

  • Sex Crimes

  • Child Exploitation

  • Defamation

  • Specialized Advice

  • Privacy

  • Intellectual Property

  • Indiscriminate Weapons

  • Hate

  • Self-Harm

  • Sexual Content

  • Elections

Custom categories

Use Add custom category to define additional content rules for your workflow. Each custom category requires:

  • Category code

  • Name

  • Description

Write descriptions that clearly define what should trigger the category.

Hallucination guard

A Hallucination guard compares an AI-generated answer with source material to identify claims that are unsupported by, or disagree with, that material.

Use it after an LLM node and before an action that sends or uses the answer.

Answer

Required. Provide the generated answer you want to check. Typically, this is the text output from a preceding LLM node.

Source context

Required. Provide the source material the answer should be based on. Examples include:

  • Relevant knowledge-base content.

  • Retrieved documents.

  • Product information or policy text.

  • Data returned by an integration.

Supply the relevant source content, rather than the generated answer itself. The guard uses this material as the reference for its check.

Question

Optional. Provide the original question that the answer responds to. This gives the guard additional context when evaluating the answer.

Sensitivity

Sensitivity controls how strictly the guard evaluates the answer against the source context. Available settings are:

  • Service default — uses Permissive sensitivity.

  • Restrictive

  • Cautious

  • Permissive

  • None

If the check is too strict for your use case, try a more lenient sensitivity. To disable the check without removing the node or changing your workflow layout, select None.

Branching

Guard nodes have two outgoing branches. Connect these to define what happens after the check.

Branch
Next step

Not triggered

Continue along the path for content that does not trigger the guard.

Triggered

Follow an alternative path, such as flagging a ticket for review.

If the selected branch is not connected to another node, the flow ends at the Guard node.

A triggered result does not automatically send a fallback message or create a review task. Configure the Triggered branch with the actions you want the workflow to take.

Example: Check a support reply before sending it

In this example, an LLM generates a support reply using knowledge-base content. A Hallucination guard checks the answer before it is sent to the customer.

1

Retrieve the source material

Retrieve the knowledge-base content relevant to the customer’s question.

2

Generate an answer

Use an LLM node to generate a reply based on that source material.

3

Configure the Hallucination guard

Add a Guard node after the LLM node and select Hallucination guard.

Field
Value to use

Answer

The generated reply from the LLM node.

Source context

The knowledge-base content used to generate the reply.

Question

The customer’s original question (optional).

Sensitivity

The setting appropriate for your use case.

4

Connect the branches

  • Not triggered: Continue to the action that sends the reply.

  • Triggered: Flag the ticket for review.

Place the action that sends the reply after the Guard node so the check happens before the customer receives it.

Node outputs

Guard results are available to downstream nodes. The available fields depend on the selected guard type.

Guard type
Output keys

Common outputs

guardType, triggered, blocking

Prompt guard

ok, reason, confidenceScore

Content guard

categories, safetyLevel

Hallucination guard

reason, level, disagreementClaims

Type-specific details may vary depending on the result returned by the guard service.

Related documentation

PreviousSwitch Nodes
NextTransformer Nodes

Last updated

Was this helpful?