Gen3 Admin Helper Settings
- Open GT Helper Settings in the Control Panel sidebar (Super Admin only).
- On the Configuration tab, enable GT Helper for the tenant app and/or Control Panel, configure Internet search, and set optional helper-specific model overrides.
- On the System prompt sources tab, review each prompt layer by scope, edit content as needed, and use Restore default to revert individual layers to factory text.
- Click Save helper settings or Save system prompt sources on the active tab before leaving the page.
- Confirm deployment-wide default models on Models when helper LLM or web search overrides are left unset.
GT Helper is the integrated ? help shelf in both the tenant app and Control Panel. These settings control whether users can open the shelf, which models power Ask GT Helper chat and optional internet search, and what system-prompt layers the deployment injects for auditability.
Misconfigured helper models produce 503 errors in the help shelf until a valid chat-capable default exists. System prompt customization affects every helper turn—keep overrides intentional and within character limits.
The page opens with summary stat cards (GT Helper on/off per app, resolved helper LLM, internet search posture) and two tabs:
| Tab | Purpose |
|---|---|
| Configuration | Enablement, internet search, helper LLM and web search model overrides |
| System prompt sources | View, edit, and restore factory prompt layers per scope |
Deep-link to a tab with ?tab=configuration or ?tab=system-prompt-sources.
Settings on this tab apply to the integrated ? help shelf in both the tenant app and Control Panel.
| Checkbox | Effect |
|---|---|
| Tenant app | Users can open GT Helper from the tenant shell ? bubble |
| Control Panel | Operators can open GT Helper from the Control Panel ? bubble |
Disabling a surface hides the help shelf there without removing wiki articles from the Instructions drawer.
- Internet search enabled for GT Helper — when on, helper chat may call the configured web search model during Ask GT Helper turns (subject to model availability and deployment policy).
- Helper web search model — optional override; leave unset to inherit the deployment default web search model from Models → Default Models.
| Field | Behavior |
|---|---|
| Helper LLM model | Optional override for helper chat. Leave unset to use default LLM from Models → Default Models (defaultInstructionsHelperModelId). |
| Helper web search model | Active only when internet search is enabled. Leave unset to inherit default web search model. |
Validation rules:
- Helper LLM must be a chat-capable enabled model, or you must have a valid deployment default LLM.
- Helper web search model must support the web search intent when explicitly selected.
- Saving is blocked when required models are missing or invalid.
Deployment-wide defaults remain authoritative on Models; this page only overrides helper-specific selections.
Every system prompt layer is visible to Super Admins. Layers are grouped by scope:
| Scope label | Typical content |
|---|---|
| Shared | Instructions common to tenant and Control Panel helper |
| Tenant app | Tenant-specific helper behavior and tone |
| Control Panel | Operator helper behavior and tone |
| Per request | Runtime context (wiki excerpts, route, workspace surface) appended after customized prefix text |
For each source you can:
- View current content (factory default or customized)
- Edit text in the source textarea (character count shown against the deployment max, typically 16,384 characters per source)
- Restore default for that layer only, reverting to GT AI OS factory text
- Save system prompt sources to persist overrides across the deployment
Customized layers show a Customized badge; factory text shows Factory default. Runtime layers may still append live values (current page path, wiki excerpts, QuickStart step context) after any customized prefix.
The deployment context source is the primary place to document deployment-specific operator guidance that should appear in helper inference.
| Page | Relationship |
|---|---|
| Models | Default LLM, default GT Helper model, and default web search model |
| GT Helper | End-user and operator help shelf behavior, retrieval, and chat |
| CTP Helper Usage | Token and thread observability for Control Panel helper |
| QuickStart | GT Helper settings step confirms helper posture during onboarding |
| Instructions Settings | Optional external instructions link surfaced in both apps (separate from helper system prompts) |
- Set Models defaults first, then refine helper overrides only when the help shelf needs a different model than general chat.
- Keep system prompt edits concise and factual; oversized layers are rejected at save time.
- Use Restore default per layer rather than deleting text manually when rolling back experiments.
- Review CTP Helper Usage after enabling internet search to monitor token burn.