Data Security

Data Processing Principles

Forwarded without retaining content: The request and response bodies are used to complete the current call and for troubleshooting, and are not used for model training;
Call logs retain only the metadata required for billing (model, token count, time, status), and can be queried in the console by yourself;
End-to-end HTTPS encrypted transmission, with keys stored in a masked form.

Key governance: split tokens by project, set quotas and expiration times, and revoke them upon departure;
Least privilege: use model groups to restrict the scope of models a token can access;
Sensitive data: before sending, mask or replace placeholder values for personal privacy fields (such as phone numbers and ID numbers), then fill them back in after getting the result;
Audit: regularly export logs to verify usage and call sources.

Data terms vary among upstream model vendors. For enterprise scenarios involving sensitive data:

Prefer model families that explicitly promise “API data is not used for training”;
When industry compliance is involved (finance, healthcare, etc.), first run a small-scale pilot and evaluate the usability and risks of the output;
For special compliance requirements (private deployment, dedicated lines, etc.), please contact customer service to discuss a solution.