Magisterium AI

Mga Developer
Ibahagi:

Safety Settings

Magisterium AI API offers user-customizable safety settings to moderate the incoming requests to the API. This moderation is currently limited to the input, not the output.

Safety Settings Behavior

If one of the safety settings is triggered, the response will either:

  • contain fallback response if response is set to true in the safety settings (defaults to true), or
  • set finish_reason to content_filter and outputs a blank text.

Example

Include safety_settings in your request:

json
{
  "model": "magisterium-1",
  "messages": [
    // your messages here
  ],
  "safety_settings": {
    "CATEGORY_NON_CATHOLIC": {
      "threshold": "BLOCK_ALL",
      "response": true
    }
  },
  // other fields ...
}

Categories and Thresholds

Caveat: CATEGORY_NON_CATHOLIC is deprecated and is no longer enforced as an API-level content filter. Instead, the assistant is instructed in its system prompt to decline all queries outside Catholicism, religion, faith, or morals. This behavior applies regardless of the configured threshold, so setting it to OFF does not disable the refusal.

The following safety categories and thresholds are supported:

CategoryDescriptionThresholdsDefault
CATEGORY_NON_CATHOLICQueries entirely unrelated to Catholicism, religion, faith and morals.BLOCK_ALL, OFFBLOCK_ALL