Model Profiles & Configuration Templates

How this fits together

There are two related concepts, managed in two different places:

Model Profiles and Configuration Templates are created and managed centrally under the Configuration library tab on the AI Agent list page.
Each individual agent then selects one of each on its Configuration (Template) tab inside AI Configuration (see AI Agent).

Configuration library

On the AI Agent page, switch from the Agents tab to the Configuration library tab. This shows two side-by-side panels:

Model profiles — "Manage the reusable model provider and generation defaults." Table columns: Name, Model, Provider, Actions. New profile button to create one.
Config templates — "Store reusable behavior, language, and style defaults." Table columns: Name, Archetype, Language, Actions. New template button to create one.

Recommended annotation: ① arrow to the Configuration library tab itself (top, to show how to get here from Agents). ② box the New profile button on the left panel. ③ box the New template button on the right panel.

Creating a Model Profile

Click New profile to open Create model profile:

Field	Description	Example shown
Name	Profile display name.	—
Provider	The LLM provider.	`openai`
Model name	Which model to use.	`gpt-5-nano`
Model version lock	Pin to a specific model version.	`gpt-5-nano`
Temperature	Response randomness. Some models only support a fixed value — the UI shows a note like "This model only supports the default temperature value of 1." when that applies.	`1`
Top P	Nucleus sampling parameter.	`1`
Frequency penalty	Reduces repetition of the same tokens.	`0`
Presence penalty	Encourages introducing new topics.	`0`
Max output tokens	Hard cap on response length for this profile.	`2000`
Fallback profile ID	Optional — a profile to fall back to if this one fails.	(empty)
Reasoning mode	Depth of reasoning the model applies.	`minimal`
Latency priority	Trade-off between speed and quality.	`balanced`
Cost tier	Cost tier for this profile.	`standard`

Recommended annotation: group with a bracket + label rather than numbering all 12 fields individually (too dense for arrows): "① Identity" over Name/Provider/Model name/Model version lock, "② Generation tuning" over Temperature/Top P/Frequency/Presence penalty, "③ Limits & fallback" over Max output tokens/Fallback profile ID, "④ Behavior" over Reasoning mode/Latency priority/Cost tier. This grouping pattern is more useful here than sequential numbers since the form isn't filled in a strict linear order.

Note: Max output tokens here is a per-profile ceiling. It's separate from the Authority Limit Level cap (level × 1000) set per agent in Governance — the two are not the same setting, and the agent is bound by whichever is more restrictive in practice.

Creating a Configuration Template

Click New template to open Create config template:

Field	Description	Example shown
Name	Template display name.	—
Archetype	The agent's overall persona archetype.	`professional`
Persuasion style	How the agent argues/persuades.	`consultative`
Language style	Formality of language.	`formal`
Verbosity level	How much detail responses include by default.	`low`
Primary language	The template's main response language.	`id`
Fallback language	Used if the primary language can't be applied.	`en`
Supported languages	Checkboxes: `id`, `en`, `zh`, `ja`, `ko`, `ms`, `th`, `vi`.	`id`, `en` checked
Translation confidence threshold	Minimum confidence (0–1) before auto-translating.	`0.8`
Version	Template version number.	`1`
Technical depth level	Slider — how technical responses can get.	`3`
Warmth score	Slider — baseline warmth in tone.	`70`
Empathy score	Slider — baseline empathy in tone.	`80`
Analytical score	Slider — how analytical/data-driven responses are.	`60`
Assertiveness score	Slider — how assertive/directive responses are.	`50`
Adaptability score	Slider — how much the agent adjusts its style to match the customer's tone.	`75`
Active voice required	Toggle — require active voice over passive voice in responses.	On
Allow emoji	Toggle — allow emoji use in responses.	On
Avoid repetition	Toggle — actively avoid repeating phrasing across a conversation.	On
Avoid metaphor	Toggle — actively avoid figurative language.	Off
Avoid marketing exaggeration	Toggle — actively avoid hype/exaggerated claims.	On
Allow auto detect	Toggle — allow the agent to auto-detect the customer's language rather than strictly using Primary language.	On

Selecting a profile and template for an agent

Inside an individual agent's AI Configuration, open the Configuration (Template) tab to select which Model Profile and Configuration Template that agent uses — see AI Agent for the tab overview.

Note: despite the name, Configuration Template sets the agent's global personality/tone baseline (archetype, language style, warmth/empathy/analytical/assertiveness scores) — the agent-specific system prompt on top of that baseline is written separately in the Behaviour tab.

Operational guidance

Create a small number of well-tuned Model Profiles (e.g., one "fast/cheap" and one "high-quality") and Config Templates (e.g., one per archetype your business needs — professional, friendly, casual) rather than a unique profile per agent — this keeps behavior predictable and easy to audit.
Frequency penalty, Presence penalty, and Top P are advanced generation controls — leave them at default unless you have a specific tuning reason, since small changes here can noticeably shift response quality.
The Warmth, Empathy, Analytical, and Assertiveness scores in a Config Template set the baseline personality — use Emotion Matrix rules on individual agents for situational adjustments (e.g., extra warmth specifically when a customer is angry) rather than trying to encode every scenario into the template itself.
Set a realistic Translation confidence threshold — too low risks mistranslated responses going to customers; too high may cause the agent to fall back to the fallback language more often than needed.
The style toggles (Active voice required, Avoid repetition, Avoid metaphor, Avoid marketing exaggeration) are useful guardrails for a professional/support-oriented archetype — turn Avoid marketing exaggeration on for any customer-support-facing template to keep the agent from overpromising.
Allow auto detect lets the agent respond in whatever language the customer writes in, even if it differs from Primary language — turn this off if you need strict language consistency (e.g., a template meant only for an Indonesian-speaking support line).
A high Adaptability score makes the agent mirror the customer's tone more — pair this thoughtfully with Emotion Matrix rules, since both mechanisms adjust tone and could compound in unexpected ways if tuned too aggressively together.