Draft — company details, retention period and legal review pending.

Data Policy

Last updated: [date]

Prompts and completions

We do not store prompts or completions. They are processed in memory only to serve the request.
Prompts and completions are not written to our logs. Our serving software logs request metadata only (see below), and request logging in the inference engine is turned off.
Recent prompt prefixes may stay in GPU memory as a short-lived prefix cache to speed up follow-up requests. This cache is never written to disk and is discarded as it is overwritten or when the server restarts.
We do not use prompts or completions to train or improve any model.

For billing, capacity planning and abuse prevention we record, per request: time, request id, API key id, model, the serving backend, status, latency, and token counts (prompt, cached prompt, completion).
Request metadata is kept for [N] days and then deleted.

Requests are processed on GPU servers rented from cloud providers (Modal and RunPod) in the United States. These providers run our software; they do not receive prompts or completions in their logs from us.