DEVELOPER DOCUMENTATION

Privacy & call data

See where your prompt goes, what we keep, and how to limit what a provider receives.

For zero data retention routing, send "zdr": true on every request. See the ZDR integration guide for copyable examples. Redaction and ZDR are separate controls.

Where your prompt goes

GPU Router’s gateway authenticates the request, applies optional redaction and chooses a route. Fusion routing also reads the request on the gateway to pick a model. We do not store ordinary prompt or response bodies, and we do not use your prompts or outputs to train models or for advertising or profiling.
The serving provider receives the request from the selected supplier route and processes your prompt and output under its own policies. On Standard routing, those policies can allow logging, retention or training. See how marketplace offers reach a provider.
The response returns through the same API. We record model, token counts, cost, timing and outcome for billing and operations.

Activity and receipts do not name the serving provider, so you cannot pick a Standard route by its data policy. When a prompt must not be retained, send "zdr": true: routing is limited to zero data retention routes, and the request fails rather than falling back to Standard. Otherwise, assume a Standard provider may keep what you send.

Redact or block

Privacy processing is optional and runs locally in the gateway. Set mode: "redact" to substitute stable placeholders within one request, or mode: "reject" to block a request containing detected identifiers before credit reservation or provider dispatch.

{
  "model": "your-model-id",
  "messages": [
    {
      "role": "user",
      "content": "Contact alice@example.com about Project Cedar."
    }
  ],
  "max_tokens": 64,
  "privacy": {
    "mode": "redact",
    "terms": [
      "Project Cedar"
    ]
  }
}

Patterns cover email addresses, common formatted phone numbers, formatted US Social Security numbers, checksum-valid payment-card numbers and selected API-key formats. Custom terms are literal, case-insensitive matches: up to 100 terms of 2–200 characters. Messages, message names, the user field and stop strings are inspected. The privacy policy and custom terms are removed before forwarding.

The playground offers the same redact/block controls. Receipts contain only category counts; responses include X-GPU-Router-Redacted-Count. Rejection returns HTTP 400 with privacy_violation. Oversized prepared requests or excessive matches are rejected.

Pattern matching can miss names, addresses, context, unusual formats and other identifying information. It can also redact innocent text. Providers still receive the remaining prompt. This feature does not guarantee anonymity or zero retention. For content that must not be retained, also send "zdr": true.

Call data and retention

Ordinary inference receipts store model, provider, tokens, cost, outcome, timing and redaction counts. The gateway does not persist ordinary prompt or response bodies in its usage ledger. Upstream failure logs omit response details. Activity exports only the loaded request metadata as JSONL; maintain any full prompt/output archive in your own application under your own retention policy.

Deferred requests require persisted content. Their prepared inputs and results are encrypted at rest with AES-256-GCM. Inputs are cleared when work finishes, expires or is cancelled. Results are downloadable for 24 hours after completion and removed by the worker afterward. Batch metadata is removed seven days after its deadline. Gateway operators control the encryption key; backups follow the operator’s separate retention policy. Provider retention is independent.