DocsReference
Glossary
Terms used across the platform and this documentation, defined once.
3 min read
Terms
| Term | Definition |
|---|---|
| Data plane | The OpenAI-compatible proxy surface — completions, model discovery, routing explain. API-key auth. |
| Control plane | The REST API for orgs, keys, credentials, providers, policies, dashboard, billing. |
| BFF proxy | The same-origin server route that injects auth so the browser never holds a raw token. |
| BYOK | Bring Your Own Keys — you supply provider credentials, envelope-encrypted per org, provider, and purpose. |
| RLS | Row-Level Security — Postgres policies enforcing tenant isolation on every tenant table. |
| Provider | A model backend. 18 total: 5 native adapters, 13 OpenAI-compatible. |
| Routing decision | The router's output: chosen provider, strategy, ranking, rejected candidates, confidence, estimated cost/latency/savings. |
| Scoring mode | A routing strategy — latency, quality, balanced, coding, reasoning, vision, long_context, embeddings, or json. |
| Circuit breaker | Per-provider health tracking that pulls unhealthy providers out of routing. |
| Semantic cache | A pgvector-backed cache keyed by embedding similarity, alongside exact-match. |
| Shadow mode | Observe-only experimentation — mirrors traffic to candidates without touching production output. |
| Entitlement | A feature or limit derived from a subscription's plan. Code checks entitlements, never plan names. |
| Plan | A billing tier — Free, Business ($149/mo), or Enterprise. |
| Estimated savings | A modelled comparison against the most expensive eligible candidate for a request — never a billing figure. |
| Usage record | A per-request metering row: provider, model, latency, tokens, retries, fallover, HTTP status. |
| ADR | Architecture Decision Record — a binding, dated record of an architectural decision. |