Guardrails
Configure the agent's guardrails to moderate and protect its interactions with users.
Guardrails
The Guardrails page configures the safety filters applied to every message the agent sends and receives. Use them to block unsuitable content, protect sensitive data, and keep the agent inside the scope you intended.
Reach it through the Guardrails icon in the editor's side navigation.
Every setting saves itself and applies to the current working version.
The tabs
| Tab | What it configures |
|---|---|
| General | Turning guardrails on, AI-assisted analysis, the blocked message |
| Harmful content | Hate speech, insults, sexual content, violence and misconduct |
| Prompt attacks | Detecting prompt injection and malicious manipulation |
| Profanity | Improper or offensive language |
| Contextual grounding check | Answers with no basis in the context documents |
| Relevance check | Answers outside the conversation's context |
| Denied topics | Specific subjects the agent must not touch |
| Sensitive information | Personal data (PII): mask it or block it |
The General tab
Turn guardrails on
Enables or disables every guardrail on the agent. With it off, no filtering happens — the agent may return unsuitable or unsafe answers.
Turning guardrails on increases the agent's response time, since every message is checked against the rules you defined.
Use AI to reinforce the guardrails
With this on, a language model reviews messages and answers to apply the guardrails in the more subjective or contextual cases — the ones deterministic filters cannot settle.
This makes extra LLM calls on every interaction, increasing token consumption and the cost per conversation. It is only available with Turn guardrails on enabled.
The message for blocked prompts
The text shown to the user when a message is blocked by the guardrails. Use it to explain the agent's policies.