Guardrails

Ensure AI-generated responses are safe, appropriate, and compliant by adding guardrail scanners to your workflows.

Overview

Guardrails are scanners that evaluate user inputs and model outputs to maintain safe, responsible AI interactions. Deploy the scanners you need, then add them to individual workflows as input or output scanners.

User Input → Input Scanners → LLM Node → Output Scanners → Response

Available Scanners

Scanner	Description	Applies To
Regex	Validates prompts using user-defined regular expression patterns. Supports defining desirable (“good”) and undesirable (“bad”) patterns for fine-grained validation.	Input
Anonymize	Removes sensitive data from user prompts to maintain privacy and prevent exposure of personal information.	Input
Ban Topics	Blocks specific topics (for example, religion) from appearing in prompts to avoid sensitive or inappropriate discussions.	Input
Prompt Injection	Detects attempts to manipulate or override model behavior, protecting the LLM from malicious or crafted inputs.	Input
Toxicity	Analyzes prompts or responses for toxic or harmful language to ensure safe and respectful interactions.	Input, Output
Bias Detection	Examines model outputs for potential bias to help maintain neutrality and fairness in generated responses.	Output
Deanonymize	Replaces placeholders in model outputs with actual values to restore necessary information when needed.	Output
Relevance	Measures similarity between the user’s prompt and the model’s output and provides a relevance score to ensure responses stay contextually aligned.	Output

Deploy Guardrails

Before adding scanners to a workflow, you must deploy them. A deployed scanner is available across all your workflows on the platform.

To deploy a scanner:

1. Click Settings on the top navigation bar.

2. Click Manage Guardrails on the left menu.

3. On the Manage guardrail models page, click Deploy next to the scanner you want to deploy.

   The status changes to Deploying. Once deployed, the scanner is available when you add scanners to a workflow.

   To remove a scanner from the platform, click Undeploy next to the scanner.

Add Scanners

After deploying the required scanners, add them to a workflow. Input scanners evaluate prompts sent to the LLM node; output scanners evaluate responses returned from the LLM. Add input and output scanners separately based on your requirements.

To add a scanner:

1. In the Workflows section, open the workflow where you want to add the scanner.
2. In the left navigation pane, click Guardrails.
3. In the Input Scanners section, click Add Scanner, select the scanners you need, then click Done. The selected scanners appear in the list.
4. Click a scanner to configure its settings. Available options vary by scanner. For example, Toxicity has Threshold and End the flow if the risk score is above settings; Regex has Enter patterns to ban, End the flow if the risk score is above, and Match type settings.
5. To add more scanners, click the Add Scanner (+) icon. To remove a scanner, click the Remove Scanner (−) icon.

Test Scanners

After adding and configuring scanners, verify that they perform as expected. You can test an individual scanner or all scanners together, then adjust settings as needed.

To test guardrails:

1. On the Guardrails page, click Test.

2. In the Prompt input box, enter a prompt or select Input template to choose a template.

   

3. Click Test. Under Scores and Results, review the output.

   Field	Description
   Validity	Indicates whether the prompt meets the scanner’s criteria. For example, if it doesn’t detect any toxicity then Validity is “True`.
   Risk Score	The prompt’s risk level, calculated as: (Threshold - Scanner Score) / Threshold. For the Relevance scanner, the score is 1 if similarity is below the threshold; otherwise it is 0.
   Duration	The time taken by the scanner to process the prompt.

4. Based on the results, adjust scanner settings and retest as needed.