Skip to content

Responsible AI Task (RAI Task)

A Responsible AI Task (RAI Task) represents a request from a customer for the RAI Service to perform specific annotations on content they send, targeting particular categories of harmful content. Each RAI Task defines its applicable scope, covering two primary dimensions:

Dimensions of an RAI Task Scope

SourceType

  • Indicates the origin of the message.
  • Values:
    • Prompt: Represents input provided by the user.
    • Completion: Represents output generated by the model.

Role

  • Specifies which role’s messages the annotation task applies to.
  • Example roles include "User," "Assistant," or other custom roles defined by the customer.

Modality

Customers do not need to specify the modality (e.g., text, image, video) of a RAI Task. The RAI Service performs annotations at the conversation level, ensuring that the modality is inherently covered without requiring explicit definition.

Use Case and Benefits

  • Comprehensive Annotation: RAI Tasks enable precise annotation of harmful content categories based on customer-specific requirements.
  • Context-Aware: By operating at the conversation level, RAI Tasks leverage the full conversational context to improve the accuracy and relevance of annotations.
  • Flexible Scoping: The dimensions of SourceType and Role allow customers to tailor annotation tasks to their specific use cases, whether focusing on user input, model output, or messages from particular roles.

By defining RAI Tasks, customers can ensure that their applications align with Responsible AI principles, maintaining safety and compliance across all interactions.