AI providers, models & capabilities

Every generative step in the pipeline — script, images, voiceover, video, music — is fulfilled by an AI model belonging to an AI provider. Administrators register these and control which model serves each step, in what order, with what fallbacks. These surfaces live under Settings → AI providers / AI models / AI roles.

🎬 Manage AI providers & models

Prefer to read? Open the step-by-step transcript

1. Settings → AI providers lists every provider with a live health badge.
2. Open a provider to see its credential reference (a Key Vault secret-* name — never a raw key), auth scheme, endpoint, and whether it has an official API.
3. Settings → AI models maps each model to its provider, capability, and cost.
4. Settings → AI roles / model-selection policy decides which model serves a step, with fallbacks tried in priority order.

The pieces

• AI provider — a vendor (OpenAI, Anthropic, ElevenLabs, Vertex, Black Forest Labs, …). Carries the endpoint, auth scheme, and a Key Vault secret reference for its credential. JARAI never stores raw keys in the database — only the secret-* name.
• AI model — a specific model offered by a provider, tagged with the capability it provides (text, image, TTS, video, music…) and its cost basis.
• AI capability — the kind of work a model can do; capability icons appear throughout the Console.
• Model-selection policy — the ordered chain that resolves “which model runs this step for this account”, honouring priority and a maximum fallback depth.

Provider health

The health badge (Healthy / Degraded / Failing / Suspended) is driven by the ProviderHealthMonitor background check. When a provider is Failing, the pipeline routes around it using the next model in the selection chain; productions keep moving rather than dead-lettering.

Tip

A provider with no configured credential, or a deliberately-inactive fallback, fails cleanly at the call boundary (e.g. a 403) and the loop advances to the next candidate — this is expected behaviour, not an outage.

Adding a provider (safe sequence)

1. Store the credential first. Put the API key in Key Vault under a secret-<vendor>-apikey-<env> name. The provider record references this name, not the secret value.

2. Create the provider in Settings → AI providers — endpoint, auth scheme, the secret-* reference, and whether it exposes an official API.

3. Add its models in Settings → AI models with the correct capability and cost.

4. Wire it into the selection chain so a step actually routes to it (otherwise it’s registered but never called).

5. Verify health turns Healthy before relying on it for live productions.

Adapters must exist before distribution records

For distribution platforms and providers that need a code adapter, the adapter must be deployed and live before the record is activated — otherwise calls fail with ADAPTER_NOT_FOUND. (This is the provider-side equivalent of the pipeline’s contract-before-seeding rule.)

Cost & spend

Each model carries a cost basis; per-call cost is recorded against every step and rolls up into the P&L dashboards. Budgets and spend limits are covered under Finance & P&L.