MaaS

Overview

MaaS (Model as a Service) manages LLM endpoint access and API keys for external providers: operators configure providers and quotas; users list endpoints and create API tokens.
Ships a Go BFF and React frontend (Mod Arch pattern).

BFF: Authenticates; resolves current user (Kubernetes RBAC or forwarded token in federated mode); proxies upstream MaaS REST API (MAAS_API_URL); reads Kubernetes for gateway and tier configuration; rate limits as Kuadrant RateLimitPolicy CRDs.
Local dev without cluster: cmd/main.go supports --mock-k8s-client and --mock-http-client; GET /healthcheck for probes and contract tests.
Federated mode:
- Module Federation remote maas: ./extensions (ODH registrations), ./extension-points (host contracts).
- Main dashboard loads remoteEntry.js and mounts MaaS in the shell.
- Base path and theme follow DEPLOYMENT_MODE / STYLE_THEME and PUBLIC_PATH when embedded.

Term	Definition
ModelEndpoint	Registered LLM inference endpoint with an OpenAI-compatible API surface
APIToken	User-scoped token for a specific model endpoint
LLMProvider	External provider whose models are exposed through MaaS
UsageQuota	Rate/token limits, stored as Kuadrant `RateLimitPolicy` CRDs
Tier	Named quota level mapped to groups via ConfigMap
MaaS Gateway	Kuadrant/Envoy ingress enforcing quotas per token

Dependency	Type	Details
`packages/gen-ai`	Package	Chat uses MaaS endpoints and API tokens
Kubernetes (Kuadrant)	Kubernetes API	`RateLimitPolicy`, `TokenRateLimitPolicy` for quotas
External LLM providers	HTTP	OpenAI-compatible traffic via configured upstream URL
Main ODH Dashboard	Host application	Federated load via remote `maas`

Themes: MUI in standalone/kubeflow; PatternFly only in federated—do not assume one UI stack across modes.
Deprecated flags: Prefer --deployment-mode=<mode> over legacy --standalone-mode / --federated-platform.
MOCK_HTTP_CLIENT: Fixtures under bff/internal/integrations/maas/testdata/ do not cover all edge cases.
Kubeflow mode: Expects a Kuadrant-capable cluster; not fully mockable.