Process the request
The gateway receives the input, applies authorized routing, sends it to the chosen provider, and streams or returns the provider response.
An AI gateway must read a request to send it to a model. The important questions are what it stores, what appears in reports, which optional features keep content, and which provider receives the request.
LLMRegistry processes prompt and completion content so it can deliver the request and response. Normal usage reporting records what happened without storing that text as part of the ledger.
The gateway receives the input, applies authorized routing, sends it to the chosen provider, and streams or returns the provider response.
Operational records can include the request ID, route, model, provider, tokens, cost, timing, cache state, and status—not the prompt or answer text.
The selected model provider receives the content needed for inference. Its own privacy terms and retention rules still apply.
A feature that reuses an answer cannot work without keeping protected data for that purpose.
Both the organization policy and the route must allow caching. A normal route does not begin saving reusable answers by accident.
The cache uses protected feature vectors and keyed digests for matching instead of saving the original prompt text.
Eligible response content is stored with authenticated AES-256-GCM encryption and a limited lifetime set by policy.
Authorized users can invalidate one project or clear the organization cache when saved responses should no longer be reused.
Application keys and provider credentials should not appear in reports or normal administration screens.
Gateway controls cannot replace the terms of the company that runs the chosen model.
Choose providers carefully, keep optional content storage explicit, and use metadata for operations without turning it into a prompt log.
Get started