Skip to content

🌐 เอกสารภาษาไทยกำลังจัดทำ — เนื้อหาด้านล่างเป็นภาษาอังกฤษชั่วคราว จนกว่าจะมีการแปล. This page is not yet translated; English content is shown temporarily.

Glossary

Key terms used throughout this documentation. For the conceptual model behind them, see Key concepts and the Multi-tenancy model.

TermMeaning
OrganizationThe top tenant boundary — a customer or business unit. The isolation, billing, and SSO boundary.
ProjectA workspace within an organization. Providers, routing, guardrails, and budgets are configured per project.
GroupA set of users within a project, used for shared limits and model allow-lists.
User / ConsumerAn identity that makes requests, expressed as the tuple organization.project.user.
Consumer tupleThe three-part identity (org.project.user) that keys API keys, budgets, usage, and isolation.
Platform adminRole spanning all organizations.
Org adminRole confined to one organization.
MemberRead-only, self-service role for an individual user.
Control planeThe configuration service (backed by PostgreSQL) that is the single source of truth.
Data planeThe gateway that handles live traffic and enforces policy. Holds no configuration of its own.
ReconcileThe continuous projection of the control plane's desired state onto the data plane.
Logical model nameA stable name developers call (e.g. coding-default), mapped to a real provider model by routing.
ProviderAn upstream LLM endpoint (OpenAI-compatible) configured per project.
RoutingThe mapping of logical model names to providers and real model IDs.
BudgetA USD spending cap at organization, project, group, or user scope; the tightest applicable cap wins.
LimitA rate cap (tokens per minute) at a tenancy scope.
GuardrailA request-content control — prompt-injection detection and PII masking.
Semantic cacheA response cache keyed by meaning, so similar prompts reuse a prior answer and avoid upstream spend.
Semantic guardAn embedding-based prompt-safety classifier, tunable per project with sample prompts.
MCPModel Context Protocol — the standard for giving AI agents access to external tools.
MCP serverA backend tool endpoint registered per project and fronted by the gateway.
GateA policy stage in the request lifecycle (key-auth, guardrails, routing, budget/limits, cache).
Key-authThe gate that authenticates an API key and resolves the consumer identity.
Tenant guardThe check that a request may only reach its own tenant's resources.
IdP brokeringConnecting an organization's own identity provider so users sign in with corporate credentials.
JIT provisioningCreating a user account automatically on first successful sign-in.
Audit logThe record of every mutating administrative action, including denied attempts.
Component matrixThe set of component versions tested together for a given product release.
Build-once / promote-by-retagShipping the exact tested image digest to production by retagging, not rebuilding.
Air-gapRunning with no internet egress — images mirrored, TLS internal, identity in-cluster.

Next steps

Enterprise AI governance, on infrastructure you own.