Config

Instruction describing how to choose among the pending instances.

The model scores each candidate 1-10 against this, against what each candidate's layout says about itself (content_description), and against any user context an app's context provider supplies. A context provider is optional: a prompt that ranks on the content alone ("Prefer time-sensitive offers over evergreen content") works without one, while a prompt about the person ("Show content matching the user's interests") only differentiates if a provider is registered for com.urbanairship.ai.Usage.Companion.embeddedSelection.

How scores and candidate priorities combine into the final order.

The score the winner must reach for the model's answer to be used at all; below it, fallback decides. Null accepts any score.

Whether a change in the pending set may re-run the model and swap the instance already on screen. When false — the default — a chosen instance keeps displaying as others come and go, until it is dismissed.

Hints carried on the subject handed to the context provider. Never rendered into the prompt.