Azure AI Language PII

Use Azure AI Language PII detection with Bifrost Guardrails for configurable PII detection, blocking, and redaction.

Overview

Bifrost Enterprise supports Azure AI Language PII as a guardrail provider for detecting personally identifiable information in request and response text.

This provider is separate from Azure Content Safety. Azure Content Safety focuses on harmful content, jailbreaks, and blocklists. Azure AI Language PII focuses on PII entity recognition.

How It Works

1. You create an Azure AI Language resource.
2. You create a Bifrost guardrail provider with provider_name: "azure-pii".
3. You optionally choose PII categories, language, model version, and domain.
4. You attach the provider to one or more guardrail rules.
5. Bifrost calls Azure Language PII analysis for selected request or response text.
6. Bifrost applies the configured action.

Azure AI Language PII currently evaluates text content. It does not inspect image pixels or arbitrary binary file contents.

Capabilities

• PII detection through Azure AI Language
• Category filtering with pii_categories
• Optional phi domain for protected health information scenarios
• API key, default credential, and Entra ID authentication
• detect_only, block, and redact actions
• Bifrost-managed redaction strategies and modes

Configuration Fields

Field	Type	Required	Default	Description
endpoint	string	Yes	-	Azure AI Language endpoint. Supports env.AZURE_LANGUAGE_ENDPOINT.
auth_type	enum	No	api_key	api_key, default_credential, or entra_id.
api_key	string	Conditional	-	Required when auth_type is api_key. Supports env.AZURE_LANGUAGE_KEY.
client_id	string	Conditional	-	Required when auth_type is entra_id.
client_secret	string	Conditional	-	Required when auth_type is entra_id.
tenant_id	string	Conditional	-	Required when auth_type is entra_id.
scopes	array	No	Azure default	OAuth scopes for token authentication.
api_version	string	No	2026-05-01	Azure Language API version.
language	string	No	en	Document language.
model_version	string	No	latest	Azure model version.
domain	enum	No	none	none or phi.
pii_categories	array	No	all categories	Azure PII categories to detect.
action	enum	No	detect_only	detect_only, block, or redact.
redaction_strategy	enum	No	replace	replace, mask, or hash. Used when action is redact.
redaction_mode	enum	No	runtime	runtime, logs_only, or runtime_reversible. Used when action is redact.
logging_opt_out	boolean	No	Azure default	Requests Azure not to log input text when supported by the service.
string_index_type	string	No	UnicodeCodePoint	Must be UnicodeCodePoint; other Azure offset modes are rejected.
timeout	integer	No	provider default	Provider execution timeout in seconds.

For the full redaction behavior matrix, see Guardrail Redaction.

Configuration

Web UI

1. Go to Guardrails > Providers.
2. Select Azure AI Language PII.
3. Click Add Configuration.
4. Enter the endpoint and authentication settings.
5. Optionally choose PII categories, language, model version, and domain.
6. Choose an action. Select Redact to enable Bifrost-managed redaction.
7. If redacting, choose the redaction strategy and mode.
8. Save the configuration and attach it to a guardrail rule.

API

curl -X POST http://localhost:8080/api/guardrails/azure-pii \
  -H "Content-Type: application/json" \
  -d '{
    "name": "azure-language-pii-redaction",
    "enabled": true,
    "config": {
      "endpoint": "env.AZURE_LANGUAGE_ENDPOINT",
      "auth_type": "api_key",
      "api_key": "env.AZURE_LANGUAGE_KEY",
      "language": "en",
      "domain": "none",
      "pii_categories": ["Email", "PhoneNumber", "USSocialSecurityNumber"],
      "action": "redact",
      "redaction_strategy": "replace",
      "redaction_mode": "runtime_reversible",
      "logging_opt_out": true
    }
  }'

config.json

{
  "guardrails_config": {
    "guardrail_providers": [
      {
        "id": 31,
        "provider_name": "azure-pii",
        "policy_name": "azure-language-pii-redaction",
        "enabled": true,
        "timeout": 10,
        "config": {
          "endpoint": "env.AZURE_LANGUAGE_ENDPOINT",
          "auth_type": "api_key",
          "api_key": "env.AZURE_LANGUAGE_KEY",
          "language": "en",
          "domain": "none",
          "pii_categories": ["Email", "PhoneNumber", "USSocialSecurityNumber"],
          "action": "redact",
          "redaction_strategy": "replace",
          "redaction_mode": "runtime_reversible",
          "logging_opt_out": true
        }
      }
    ]
  }
}

Helm

bifrost:
  guardrails:
    providers:
      - id: 31
        provider_name: "azure-pii"
        policy_name: "azure-language-pii-redaction"
        enabled: true
        timeout: 10
        config:
          endpoint: "env.AZURE_LANGUAGE_ENDPOINT"
          auth_type: "api_key"
          api_key: "env.AZURE_LANGUAGE_KEY"
          language: "en"
          domain: "none"
          pii_categories:
            - "Email"
            - "PhoneNumber"
            - "USSocialSecurityNumber"
          action: "redact"
          redaction_strategy: "replace"
          redaction_mode: "runtime_reversible"
          logging_opt_out: true

Authentication Modes

auth_type	Required fields	Notes
api_key	endpoint, api_key	Simple key-based authentication.
default_credential	endpoint	Uses Azure default credentials, such as managed identity or Azure CLI credentials.
entra_id	endpoint, client_id, client_secret, tenant_id	Uses an Entra ID service principal.

Category Filters

Leave pii_categories empty to let Azure evaluate all supported categories for the selected language, model, and domain. Set pii_categories when you want a narrower policy, such as Email, PhoneNumber, or USSocialSecurityNumber.

Bifrost requires string_index_type: "UnicodeCodePoint" because Azure’s returned offsets must map correctly back to Bifrost’s text ranges. Other Azure string index modes are rejected during configuration validation.

Operational Notes

• The default action is detect_only, so set action: "redact" or action: "block" when you want enforcement.
• domain: "phi" enables Azure’s protected health information domain where supported by the service.
• logging_opt_out controls Azure service-side logging behavior; it is separate from Bifrost’s own content logging settings.
• Redaction modes affect Bifrost runtime payloads, Bifrost logs, and trace-export connectors as described in Guardrail Redaction.