3. Request and response content
The prompts, code, files and images you send, and the responses the models return, are proxied between your tool and the model provider over encrypted connections. Content is processed in memory to route, sanitise and meter the request. It is not stored in a database, not used to build profiles, and not used to train models by us.
Operational logs may transiently capture fragments of content when a request errors, so that faults can be diagnosed and abuse investigated. These logs are short-lived and access to them is restricted to the people operating the service.
Model providers process content under their own terms in order to generate a response. We select providers whose API terms do not permit training on customer content submitted through the API.