Zero Data Retention

Zero Data Retention (ZDR) means that an AI model provider is not permitted to store inference data for any period of time. Providers that do not retain inference data are also unable to train on that inference data.

TypeDuck can route supported AI requests through providers or endpoints that advertise ZDR-compatible handling where available. Some providers may not train on customer data but may still retain it for abuse monitoring, legal, or operational reasons. ZDR and training controls are related but separate data-handling policies.

TypeDuck tracks model-provider and endpoint data-handling policies so eligible requests can be routed according to your configured retention requirements. A provider's general policy may differ from the policy for a specific endpoint, deployment, or enterprise route.

TypeDuck may enforce ZDR independently for different model groups. This is useful when your organization needs stricter retention controls for certain model families while allowing broader routing for other workloads.

TypeDuck can represent ZDR enforcement through separate guardrail fields for each model group. These controls allow retention requirements to be applied to different workspaces, API keys, users, or automation flows.

ZDR can also be requested for a specific inference request using a provider preference such as zdr. Request-level ZDR works as an OR with workspace-wide and guardrail settings: if any setting requires ZDR, ZDR routing is applied.

{
  "model": "gpt-4",
  "messages": [...],
  "provider": {
    "zdr": true
  }
}

Some model endpoints provide implicit in-memory caching of repeated prompt segments inside the provider datacenter. TypeDuck treats transient in-memory prompt caching as distinct from persistent data retention when the provider policy states that cached data is not stored for later use or training.

ZDR controls are designed to reduce retention by AI model providers during inference. TypeDuck may still retain application data needed to provide the product, such as account records, workspace configuration, chats, uploaded files, connector grants, audit metadata, and outputs, as described in the TypeDuck Privacy Policy.