Experts (Preset AI Roles)
An Expert freezes "one system prompt + one model + one set of runtime parameters" into a reusable identity. Once configured, users pick it at the top of the chat box and the AI answers as a support agent, a SQL engineer, a finance reviewer — instead of everyone restating the same brief every time.
Experts are maintained under AI Manager → Expert, and drag-sorting them sets the order they appear in the chat box.
What it solves
| Without experts | With experts |
|---|---|
| Everyone writes their own prompt; quality varies and results are unstable | Prompts are maintained centrally, so everyone works from the same standard |
| Every new scenario means restating the background | Switching expert switches the role and the tone |
| Using a stronger model for one scenario means changing the global default | Each expert binds its own model, so the expensive one is used only where it matters |
| Temperature and thinking mode are global settings | Parameters travel with the expert: low temperature for precise work, high for creative |
Configuration
| Field | Meaning |
|---|---|
| Name | The expert's name in the chat box |
| Enabled | Disabled experts do not appear in the selector |
| LLM | The model this expert uses — leave blank to use the default chat model |
| Prompt | The system prompt: who this expert is, what it is responsible for, how it should answer |
| Prompt Handler | Optional, rewrites the prompt at runtime — see below |
| Hint List | Starter questions shown once the expert is selected, so users are not left wondering what to ask |
| Description | What the expert is for |
| Expert Config | Runtime parameters as JSON: temperature, top_p, strictTools (strict tool calling) and thinking (reasoning mode, for models like DeepSeek-R1) |
Dynamic prompts
A prompt fixed in the configuration often is not enough — the same expert facing different tenants or differently privileged users needs different context. Implement EruptPromptHandler to rewrite the prompt before each conversation; register it as a Spring Bean and it becomes selectable in the expert form:
@Component
public class TenantPromptHandler implements EruptPromptHandler {
@Override
public String name() {
return "Tenant context";
}
@Override
public String handle(String prompt) {
// append runtime context after the prompt the administrator wrote
return prompt + "\nCurrent tenant: " + TenantContext.name()
+ "\nToday: " + LocalDate.now();
}
}Typical uses: injecting the current user and roles, the tenant and its data scope, a date-sensitive note, or knowledge-base retrieval results.
Layers of prompting
The global system-prompt applies to every conversation, prompt engineering adds application- and request-level injection points, and an expert is the identity-level prompt the user picks. All of them stack into the system prompt for the turn.
Contributors
YuePeng
Claude Opus 5 (1M context)