# NeMo Guardrails

NeMo guardrails use NVIDIA NeMo Guardrails to check agent inputs and outputs. One NeMo guardrail can combine several checks — called rails — such as forbidden patterns, a written company policy, and NVIDIA NemoGuard safety models.

## Before You Start

NeMo guardrails are off by default. Your platform administrator must:

- Set `FF_NEMO_GUARDRAILS_ENABLED` and `NEXT_PUBLIC_FF_NEMO_GUARDRAILS_ENABLED` to `true` on the Bot Builder service. The **NeMo** tab then appears on the Guardrails page.
- Run guardrails in direct mode (`FF_GUARDRAIL_MODE=direct`). NeMo guardrails cannot be created or used in proxy mode.
- For the **Content Safety** and **Topic Safety** rails, register an NVIDIA NemoGuard model in **Control Center → LLMs** and enable it for the workspace under [Integrations → LLMs](../integrations/llm_models.html). Only enabled models whose provider or identifier contains “nvidia” are offered.

## Creating a NeMo Guardrail

On the **NeMo** tab, click **+** (**Create NeMo Guardrail**). Turn on at least one rail. Each rail’s fields appear when you switch it on.

[![Create NeMo Guardrail form — section navigation, Basic Information, and the Self-Check and Regex Check rails](../../images/nemo-guardrail-form.png)](../../images/nemo-guardrail-form.png)

**Basic Information**

- **Name** — lowercase letters, numbers, and hyphens, up to 50 characters. Cannot be changed after the guardrail is created. Required.
- **Tags** and **Labels** — metadata for organisation and governance. Press Enter after each one.

**Self-Check**

Checks messages against a written policy, using the model of the agent or Guardrail component the guardrail is attached to.

- **Enable Input** — checks the user’s message before the agent responds. **Self-Check Input Prompt** holds your policy for user messages; it starts with a default policy you can edit.
- **Enable Output** — checks the agent’s response. **Self-Check Output Prompt** holds your policy for agent messages.

Write the policy as a list of rules. The platform adds the message and a yes/no question to the end of your prompt.

**Regex Check**

Blocks messages that match any of your patterns. No model is needed.

- **Input Patterns** — patterns checked against user messages, for example `(?i)ssn|social.?security`.
- **Output Patterns** — patterns checked against agent responses, for example `(?i)api[_-]?key`.

Press Enter after each pattern. At least one pattern is required for each direction you turn on.

**Content Safety**

Blocks unsafe content using an NVIDIA NemoGuard content safety model. Turn on **Enable Input**, **Enable Output**, or both, and choose the **Content Safety Model**. The model checks a fixed set of categories, including violence, sexual content, hate, harassment, self-harm, PII and privacy, fraud, malware, and unauthorised advice.

**Topic Safety**

Keeps users on allowed topics. Applies to user messages only.

- **Topic Safety Model** — an NVIDIA NemoGuard topic control model. Required.
- **Allowed / Blocked Topics** — one line per rule describing what is on-topic and off-topic. For example: “Allow questions about our product, pricing, and support.”, “Block politics, medical advice, and legal advice.” Required.

**Blocked Messages**

- **Blocked Input Message** — shown to the user when their message is blocked. Up to 500 characters. Required.
- **Blocked Output Message** — shown when the agent’s response is blocked. Up to 500 characters. Required.

**Error Handling Config**

- **Strategy** — **Fallback**.
- **Fallback Response** — returned when the guardrail cannot run, for example when a safety model is unavailable. It is not used when a rail blocks a message. Required.

When a NeMo guardrail cannot run, the message is not passed through — the user sees the fallback response.

## How the Rails Run

Rails run in this order, and checking stops at the first rail that blocks: Regex, Self-Check, Content Safety, Topic Safety. A blocked user message returns the **Blocked Input Message**; a blocked response returns the **Blocked Output Message**.

## Using a NeMo Guardrail

Apply a NeMo guardrail the same way as other guardrails:

- **Agentic Flow** — in the **Guardrail** component, or in the **Input Guardrails** and **Output Guardrails** settings of an **Agent** component. The Self-Check rails use the model selected on that component.
- **AI Coworkers** — in the **Input guardrail** and **Output guardrail** dropdowns on the [Governance](../../coworkers/governance.html) tab.

## Managing NeMo Guardrails

Each save creates a new version, and agents use the latest version. Click a card to edit it, or use the card menu to **Share** it. Deleting a NeMo guardrail disables all of its versions. NeMo guardrails can also be exported and imported in bulk with the other guardrail types.