Privacy & data policy
Service operator
Retrivium is operated by IE Aleksandr Golubev, Tbilisi, Georgia.
For questions about this policy and data handling, contact golubev.cto@gmail.com.
Informational website
This site provides information about Retrivium, GPU infrastructure and work to connect its pool to OpenRouter as an inference provider. It has no application forms, user accounts, payments or API-key registration.
Website storage and analytics
No advertising trackers or third-party analytics are included. Fonts are hosted with the site. The article catalogue can read search and category parameters from the page URL.
Website hosting logs
The website hosting provider may process technical information such as IP addresses and request times. Retention depends on the hosting configuration used for publication. These website access logs are separate from inference API requests and do not establish a retention policy for inference content.
Production inference API data policy
We process your inference data only to deliver the requested result. We do not retain it, analyse its content, use it for training or review it manually. These rules apply to ordinary and streaming API responses.
Processing only to serve your request
Retrivium processes inference inputs and outputs solely to authenticate, route, validate and execute the requested inference and return its result. This includes messages, system instructions, tool definitions and arguments, schemas, generated text and reasoning. We do not execute client tools; their definitions and results are model input.
No training, content analysis or human review
We do not use inference inputs, outputs, reasoning or tool data for training, fine-tuning, evaluations, benchmarking, product improvement, analytics, profiling, advertising or dataset creation. We do not manually review this content, sell it, or disclose it for those purposes. Operational performance tests use synthetic or operator-authored test data, never client inference content.
Zero persistent retention of inference data
We do not persist inference content, including token sequences derived from it, to disk, databases, object storage, logs, traces, monitoring systems or backups. We keep no persistent records of individual inference requests: no request IDs, timestamps, model selections, token counts, status codes, network addresses or credential identifiers associated with a particular inference request. There is no content-retention exception for debugging, quality review or abuse analysis.
Temporary processing in RAM and GPU memory
Inputs, outputs and temporary token representations exist in RAM and GPU memory only as needed to process and deliver a request. Active-request state is released when completion, cancellation or failure is handled. Inference data is not written to swap, disk-backed caches or core dumps. Cross-request prefix caching and persistent KV caching are disabled on production serving endpoints. Memory allocation may be reused; this is not a promise of immediate cryptographic or physical erasure of every freed memory block.
Aggregate operational counters only
Routing uses temporary token counts, active-request counts and KV-memory occupancy. We may retain aggregate service or model totals, such as total input and output tokens or completed-request counts. These contain no content, request identifiers, per-request event history or client profiles. Per-request usage is returned to the caller; Retrivium does not store that response as an individual billing record.
Infrastructure and processing locations
Production inference runs exclusively on hardware owned and operated by Retrivium. For endpoints designated EU, Retrivium guarantees that request routing, input processing and model inference take place exclusively within European Union member states. Prompts, responses, reasoning and tool data processed by Retrivium remain within the EU; we do not transfer them to infrastructure outside the EU. Load balancing, retries and failover for EU endpoints use only EU machines. If EU capacity is unavailable, the request is rejected rather than rerouted outside the EU. Actual serving countries are published for each endpoint. The operator’s registered location in Tbilisi, Georgia does not change the processing region of an EU endpoint. This commitment covers Retrivium processing; OpenRouter, client applications and client-controlled external tools have their own policies.
Transport and access controls
Public API traffic uses HTTPS. Inference traffic between machines is encrypted and authenticated. Production machines must pass ownership, access-control, no-logging, no-swap and no-core-dump checks before activation. Inference content is excluded from access logging, persistent application output and backups. Administrative access is restricted to authorised operators and is not used to inspect client inference content.
OpenRouter and client-controlled services
This policy covers Retrivium processing. OpenRouter applies its own privacy, routing, billing and safety policies before and after calls to Retrivium. Client applications and services invoked by client tools also have their own policies. Their independent retention settings are not controlled by Retrivium. Any OpenRouter ZDR label is assigned by OpenRouter; this page does not claim approval, certification or provider acceptance.
Requests about your data
Contact the operator using the email above for privacy questions or requests concerning data rights. Because inference content and individual request records are not retained, we cannot retrieve a past prompt, response or conversation. API credentials, configuration and voluntarily sent support correspondence are separate service-administration data, not inference history. Do not send prompts, responses or secrets in support email; correspondence is used only to handle your enquiry and any applicable legal obligations.
Changes to this policy
The date above identifies this version. Changes apply prospectively and do not authorise reuse of earlier inference data. We do not silently enable content logging, retention or training. If a future service requires different data practices, its policy and terms must be disclosed before it receives client traffic.