Setting Up GenAI Metrics

Learn how to configure built-in and custom metrics that evaluate your GenAI model’s prompts and outputs for safety, compliance, and quality.

Accessing the Model Metrics Page

Navigate to Model Overview → Model Management → Metrics.

Metrics Navigation

This opens the Model Metrics configuration screen:

Metrics Screen

If your model includes default rules, they will appear but cannot be modified here:

Default Rules

Adding a New Metric

Click + Metric to open the metric creation form:

Add Metric Form

You must provide:

  1. Metric Name
  2. Metric Type

Supported metric types include:

  • Toxicity
  • PII
  • Sensitive Data
  • Prompt Injection
  • Hallucination
  • Keyword
  • Regex

Metric Types & Configuration

Toxicity

Set a threshold using the slider or manual value input.

Toxicity Config

Applies to: Prompt, Response, or Both.


PII

Choose which PII entities the metric should detect.

Loom demo: https://www.loom.com/embed/22f017056bad46a8a0d64eea1ed157f6

Applies to: Prompt and Response.


Sensitive Data

Add custom sensitive terms and flag them accordingly.

Loom demo:
https://www.loom.com/embed/56575adfd48d4ccab9d9b54a496abf90

Applies to: Prompt only.


Prompt Injection

Detects attempts to manipulate or override system instructions.

Prompt Injection

Applies to: Prompt only.


Hallucination

Detects unsupported or fabricated model responses.

Hallucination

Applies to: Response only.


Keyword

Provide a list of keywords you want to detect.

Keyword

Applies to: Prompt, Response, or Both.


Regex

Define regex-based pattern detection.

Regex

Applies to: Prompt, Response, or Both.


Reviewing & Managing Metrics

All newly added metrics appear in the list:

Loom walkthrough:
https://www.loom.com/embed/dcf4521aa41e4bf9bbfb94029fe53d9d

New metrics remain enabled until saved:

Enabled Metrics

You may edit or remove metrics at any time:

Loom demo:
https://www.loom.com/embed/7019ea04b7114b4f8f152a2bd1a42930

Once everything is configured, click Save. After saving, each metric can be toggled on/off as needed:

Save Metrics

What’s Next

  • Run evaluation pipelines using the metrics you've configured.
  • Enable real-time guardrails to enforce these checks at inference time.
  • Explore additional guides:
    • Toxicity Detection
    • Sensitive Data Leakage
    • Prompt Injection Detection
    • Hallucination Detection