AI Control
Ce contenu n’est pas encore disponible dans votre langue.
AI Control
Section titled “AI Control”Overview
Section titled “Overview”AI Control is the dedicated page for AI preferences — the cross-environment knobs that decide how the platform uses AI:
- Model routing — the routing strategy (free-only / free-first / economy / quality), a catch-all default model, and per-need First/Fallback pins (vision, reasoning, agent, fast, long-context).
- The Jury — the best-of-N candidates and judges that compete on a turn.
- Response cache — serve byte-identical re-requests for free.
- Prompts — domain knowledge injected into each AI context (dashboard generation, widget refinement, the Sidekick chat, the dashboard consultant).
Everything here is one shared configuration, persisted to the registry and applied across all environments.
Credentials live elsewhere. Provider API keys, each provider’s enabled-model pool, and billing/admin secrets stay on Service Stack → AI — they are per-environment secrets, not preferences. AI Control links there, and back.
How it works
Section titled “How it works”When the platform makes an AI request it resolves a need (e.g. agent, vision,
reasoning) to a concrete model through the routing chain:
- Strategy filter — narrow the eligible pool to models that match the selected tier (free-only, free-first, economy, quality).
- Per-need pin — the First model for that need is tried first; if unavailable
(quarantined by a provider-supplied
Retry-After) the gateway falls over to the Fallback in the same request, then to the catch-all default. This failover is in-turn — the caller always receives an answer without retrying. - Jury (optional) — when enabled, N candidates independently answer the same prompt and a judge model scores them; the highest-scoring answer is returned.
- Response cache — identical requests (same hash of model + messages) are served from the cache without a provider round-trip, at zero cost.
Provider quarantines use only the window the provider signals (RetryInfo /
Retry-After); no invented cooldown durations are applied. A quarantine that
carries no explicit window is not applied — the request goes to the provider.
How prompts work
Section titled “How prompts work”Each context has a prompt slot. A blank slot uses the built-in default; typing a value overrides it. Clearing the field restores the default.
| Slot | Injected when |
|---|---|
| Global prompt | Prepended to every AI request |
| Dashboard generation | The AI builds a full dashboard from a schema |
| Widget refinement | The AI edits or rewrites a single widget |
| Sidekick chat | Every Sidekick conversation, as persistent context |
| Dashboard consultant | The structured business-case consultant flow |
Step-by-step
Section titled “Step-by-step”- Open AI Control (Administration → System → AI Control, next to Service Stack).
- Set the routing strategy, default model, and per-need pins under Model routing.
- Curate candidates and judges under Jury (optional).
- Toggle the Response cache and tune its TTL / max entries (optional).
- Edit any Prompt slot to steer a context; leave a slot blank for its default.
- Changes are shared across all environments — persisted to the registry.
Common mistakes
Section titled “Common mistakes”Do not treat AI output as final. Users should review generated transforms, widgets, and explanations.
Do not paste secrets into a prompt — prompts are stored in the registry config blob, not the secret store. API keys belong on Service Stack → AI.