Skip to content

Data handling and retention

No prompt, completion, image or embedding input is written to disk, logged or attached to a trace, anywhere in siema_ai. Request content exists in memory on the gateway and on the model server for the duration of the request and is gone when the response ends. This is a design constraint of the platform, not a setting.

If you send us a request_id, we can see the metadata below and nothing else.

Usage metadata, for every request that reached a model server, for 13 months. A request rejected before reaching a model server — a bad request, an empty balance, an unknown model, a rate limit, no capacity, or a gateway error checking your balance — leaves no usage record at all; see Errors for the full list of pre-routing statuses:

Field Example
Request id, organisation id, API key id 0192a1b2-…, your org, the key’s id — never the key itself or its display prefix
Endpoint, model alias, deployment id, upstream model name chat/completions, polish-pro, a deployment id, Bielik-11B-v3
Stream flag true
Input tokens, output tokens, usage source 2 010, 488, upstream (or estimated when the request was disconnected or failed before the model server reported usage, streaming or not)
Status, error code, HTTP status ok, -, 200
Time to first token, total time 2.8 s, 9.1 s
Region unknown (the serving gateway instance’s own configured region label, defaulting to unknown until it’s set)
Cost in micro-PLN 4 479
Start time, finish time timestamps

Billing keeps a separate, append-only ledger entry for every charged request — the request id, the model alias, the amount, and the time — alongside your organisation’s other billing records (top-ups, trial credit, adjustments); entries can never be updated or deleted. How long the ledger itself is kept has not been decided yet.

Console audit events (key created, member invited, and similar organisation and account changes) are kept indefinitely. Payment records are kept for the period Polish tax law requires.

Account data (email, password hash, organisation membership) is deleted on request; email support@siema-ai.pl from the owner’s address.

Everything runs in the European Union. Model inference runs on serverless GPU capacity in EU data centres; the gateway, console and databases run on a Hetzner server in Nuremberg, Germany. No request content leaves the EU. We do not use your requests to train or fine-tune any model.

Subprocessor Purpose Location
Runpod, Inc. Serverless GPU inference (model servers) EU regions
Hetzner Online GmbH Gateway, console, databases; nightly backups (Storage Box) Nuremberg and Falkenstein, Germany
Our transactional email provider (to be named before launch) Account emails: verification, invitations, password reset EU
PayU S.A. Payment processing (top-ups) Poland

This lists every subprocessor category we use or plan to use; one row is a placeholder until a vendor is chosen before launch. Changes are announced in the changelog at least 30 days before a new subprocessor handles request metadata.

  • Responses come from an AI system. If you show them to your own users, EU AI Act transparency rules may require you to say so.
  • A data processing agreement is available on request from support@siema-ai.pl.