Patronus AI

Integrate Patronus AI with Bifrost for LLM security and safety including hallucination detection, PII identification, toxicity screening, and custom evaluators.

Overview

Bifrost Enterprise supports Patronus AI as a third-party guardrail provider for evaluating LLM request and response text with Patronus evaluators.

Use it when you want evaluator-based checks such as PII detection, toxicity screening, prompt-injection checks, response quality criteria, or custom evaluators from your Patronus account.

How It Works

You’ll need a Patronus API key to authenticate with their Evaluate API - grab one from the Patronus dashboard.

1. You create a guardrail provider with provider_name: "patronus-ai" and your Patronus API key.
2. You configure one or more Patronus evaluators.
3. You attach that provider to a guardrail rule.
4. The rule decides when to run the provider and whether to evaluate input, output, or both.
5. Bifrost calls the Patronus Evaluate API at /v1/evaluate.
6. If any evaluator returns pass: false, Bifrost returns GUARDRAIL_INTERVENED.

The Patronus evaluator flow supported here is text-based: Bifrost sends selected request or response text as the evaluation input.

Capabilities

• PII Detection: Identify personally identifiable information using Patronus evaluators
• Toxicity Screening: Evaluate text for toxic or unsafe content
• Prompt Injection Checks: Use Patronus judge criteria such as patronus:prompt-injection
• Response Quality Checks: Evaluate outputs for criteria such as conciseness, helpfulness, politeness, JSON validity, code validity, or CSV validity
• Bias Checks: Use Patronus criteria for age, gender, and racial bias checks
• Custom Evaluators: Use evaluator IDs and criteria configured in your Patronus account

Configuration Fields

Field	Type	Required	Default	Description
api_key	string	Yes	-	Patronus API key. Supports env.PATRONUS_API_KEY.
base_url	string	No	https://api.patronus.ai	Custom Patronus API base URL. Bifrost appends /v1/evaluate.
evaluators	array	Yes	-	Patronus evaluator entries to run. At least one is required.
capture	enum	No	none	Controls whether Patronus stores evaluation results: none, fails-only, or all.
timeout	integer	No	30	Provider execution timeout in seconds.

Evaluator Fields

Field	Type	Required	Description
evaluator	string	Yes	Patronus evaluator name, such as pii, toxicity-perspective-api, judge, or a custom evaluator ID.
criteria	string	No	Criteria/profile name for evaluators that require one, for example patronus:is-concise.
explain_strategy	enum	No	When to include evaluator explanations: never, on-fail, on-success, or always.

Capture Modes

Captured evaluation results appear under the Traces section in the Patronus dashboard.

Value	Meaning
none	Do not capture evaluation results in Patronus.
fails-only	Capture only failed evaluator results in Patronus.
all	Capture all evaluator results in Patronus.

Explanation Response Modes

Value	Meaning
never	Do not request evaluator explanations.
on-fail	Request explanations for failed evaluator results.
on-success	Request explanations for passed evaluator results.
always	Request explanations for all evaluator results.

Built-In UI Presets

The Bifrost dashboard exposes common Patronus evaluator presets:

Preset	Evaluator	Criteria
Detect PII	pii	-
Detect Toxicity	toxicity-perspective-api	-
Prompt Injection	judge	patronus:prompt-injection
Answer Refusal	judge	patronus:answer-refusal
Is Concise	judge	patronus:is-concise
Is Helpful	judge	patronus:is-helpful
Is Polite	judge	patronus:is-polite
No Apologies	judge	patronus:no-apologies
No OpenAI Reference	judge	patronus:no-openai-reference
No Age Bias	judge	patronus:no-age-bias
No Gender Bias	judge	patronus:no-gender-bias
No Racial Bias	judge	patronus:no-racial-bias
Is JSON	judge	patronus:is-json
Is Code	judge	patronus:is-code
Is CSV	judge	patronus:is-csv

You can also select Custom evaluator and provide your own evaluator and optional criteria.

Configuration

Web UI

1. Go to Guardrails > Providers.
2. Select Patronus AI.
3. Click Add Configuration.

4. Enter a descriptive Name.
5. Set your API Key directly or through an environment variable.
6. Leave Base URL empty or use the default https://api.patronus.ai, or set a custom Patronus endpoint.
7. Add one or more evaluators.
8. Choose a Capture mode. Bifrost defaults to None.
9. Set the timeout and save the configuration.
10. Attach the configuration to a guardrail rule under Guardrails > Configuration.

API

Create the Patronus AI provider configuration directly with the management API. The Enterprise backend registers guardrail provider APIs at /api/guardrails/{provider}; the provider type is the path segment (patronus-ai), and the API assigns the configuration ID after creation.

curl -X POST http://localhost:8080/api/guardrails/patronus-ai \
  -H "Content-Type: application/json" \
  -d '{
    "name": "patronus-quality-checks",
    "enabled": true,
    "config": {
      "api_key": "env.PATRONUS_API_KEY",
      "base_url": "https://api.patronus.ai",
      "evaluators": [
        {
          "evaluator": "pii",
          "explain_strategy": "on-fail"
        },
        {
          "evaluator": "judge",
          "criteria": "patronus:is-concise",
          "explain_strategy": "on-fail"
        }
      ],
      "capture": "none",
      "timeout": 30
    }
  }'

To attach it to a rule, fetch the generated config ID with GET /api/guardrails/patronus-ai, then reference it in selectedGuardrailProfiles (in the form patronus-ai:<id>) on POST /api/guardrails/rules.

config.json

{
  "guardrails_config": {
    "guardrail_providers": [
      {
        "id": 40,
        "provider_name": "patronus-ai",
        "policy_name": "patronus-quality-checks",
        "enabled": true,
        "timeout": 30,
        "config": {
          "api_key": "env.PATRONUS_API_KEY",
          "base_url": "https://api.patronus.ai",
          "evaluators": [
            {
              "evaluator": "pii",
              "explain_strategy": "on-fail"
            },
            {
              "evaluator": "judge",
              "criteria": "patronus:is-concise",
              "explain_strategy": "on-fail"
            }
          ],
          "capture": "none"
        }
      }
    ],
    "guardrail_rules": [
      {
        "id": 401,
        "name": "patronus-openai-output",
        "description": "Run Patronus checks on OpenAI responses",
        "enabled": true,
        "cel_expression": "provider == 'openai'",
        "apply_to": "output",
        "sampling_rate": 100,
        "timeout": 30,
        "provider_config_ids": [40]
      }
    ]
  }
}

Helm

bifrost:
  guardrails:
    providers:
      - id: 40
        provider_name: "patronus-ai"
        policy_name: "patronus-quality-checks"
        enabled: true
        timeout: 30
        config:
          api_key: "env.PATRONUS_API_KEY"
          base_url: "https://api.patronus.ai"
          evaluators:
            - evaluator: "pii"
              explain_strategy: "on-fail"
            - evaluator: "judge"
              criteria: "patronus:is-concise"
              explain_strategy: "on-fail"
          capture: "none"

    rules:
      - id: 401
        name: "patronus-openai-output"
        description: "Run Patronus checks on OpenAI responses"
        enabled: true
        cel_expression: "provider == 'openai'"
        apply_to: "output"
        sampling_rate: 100
        timeout: 30
        provider_config_ids: [40]

For provider comparison and information on configuring guardrail rules and profiles, see Guardrails.