Skip to main content

NEW You.com Inference Router: Enterprise-grade inference on demand, with an uptime SLA of at least 99.9% and ZDR.

NEW Inference Router: Enterprise-grade inference with ZDR and 99.9% uptime SLA.

Learn More

Last updated:

Inference Router ZDR

ZDR stands for Zero Data Retention.

Generally speaking ZDR means a provider processes a prompt, returns a response, and doesn’t store it afterward. In other words, the provider doesn’t keep your inputs or the model’s outputs once the request is finished. The only exceptions to retention of any data relate to actual provision of the service for things like tracking for usage and payment.

Specifically ZDR means the provider operates purely in-memory: prompt and model response are processed, returned to the customer, and immediately wiped from RAM without persistent disk storage, databases, or logs.

What it takes to be ZDR compliant

Being ZDR compliant requires the provider to completely isolate the request pipeline from all persistent storage:

  • No Model Training: Prompts and responses are never used to train or fine-tune any models including the provider’s own models or services.
  • No Trust & Safety / Abuse Logging Persists: Standard API plans usually store prompts for 30 days to screen for policy violations or malware. ZDR would mean those logs are promptly reviewed, written over and do not persist.
  • No Diagnostic / Request Logging: Full text payloads are omitted from system logs; only metadata to operate the services (token counts, timestamps, status codes) is retained for billing.
  • Stateless Infrastructure: The server processes the request as an atomic, isolated transaction and immediately clears its cache.

What ZDR excludes: examples

  • Payload Hashes / Fingerprints: Creating a hash (e.g., SHA-256) of the prompt text to detect duplicate incoming requests.
  • Partial / De-identified Texts: Storing prompt templates, system instructions, or sanitized bodies.
  • Cached In-Memory Payloads on Disk: Writing prompts to persistent disk-backed prompt caches across server restarts.
  • Short term retention for abuse evaluation

If an auditor checked a ZDR system, finding actual prompt/response data saved to a persistent database or disk, even when anonymized or de-duplicated, that would violate the agreement. Our approach to ZDR ensures no part of the You.com processing activities on the Inference Router services violate these ZDR principles.

What ZDR allows: examples

Providers can still collect non-content operational metadata and control data (everything in control plane memory and DB) needed to bill you, keep infrastructure running and other operational data needed to implement services. Some examples of metadata that would be retained:

MetricAllowed under ZDR?Example
Token CountsYesinput_tokens: 450, output_tokens: 120
Request MetadataYesTimestamps, HTTP status codes (200 OK), Latency (ms)
Account IdentifiersYesorganization_id, api_key_id
Anonymized AggregatesYesTotal API calls per hour, rate limit tracking counters