Expert Templates¶

Expert Templates define which LLMs in which configuration are used for which task category. They are the central control instrument for MoE routing.

Expert Templates

Overview¶

An expert template defines:

Which LLM models are responsible for which expert category
Whether a model is required (Required) or optional (Two-Tier fallback)
Optional system prompts per category
An optional planner LLM for routing decisions
An optional judge/merger LLM for response synthesis

Template Management (`/templates`)¶

Template Cards¶

Each template is displayed as a card with:

Name and description
Category badges (which expert categories are configured)
Badge Planner (if planner LLM is set)
Badge Judge (if judge LLM is set)
Actions: Edit, Copy as Template, Delete

Creating / Editing a Template¶

Basic data:

Field	Description
Name	Unique display name
Description	Optional free text

Expert configuration (per category):

Enable category (checkbox)
Add LLMs (one or more rows):
LLM@Node: Model and inference server (format: model:tag@server)
Required: Model must be available; if missing, the request fails
Optional (Two-Tier): Used only if the required model does not respond
Optional: enable and edit a system prompt for this category

Orchestration prompts (optional):

Area	Description
Planner LLM	Model for routing decisions (overrides global PLANNER_MODEL)
Planner Prompt	System prompt for the planner LLM
Judge/Merger LLM	Model for response synthesis (overrides global JUDGE_MODEL)
Judge Prompt	System prompt for the judge LLM

Standard Expert Categories¶

Category ID	Display Name	Use Case
`general`	General	General queries
`code_reviewer`	Code Review	Code analysis and reviews
`technical_support`	Technical Support	IT issues, debugging
`data_analyst`	Data Analysis	Statistics, CSV, data
`creative_writer`	Creative Writing	Texts, ideas, storytelling
`medical_consult`	Medicine	Medical questions
`legal_advisor`	Legal	Legal questions
`math`	Mathematics	Calculations, proofs
`science`	Science	Natural sciences
`reasoning`	Reasoning	Logic, conclusions
`translation`	Translation	Language translations
`vision`	Vision	Image analysis
`financial_advisor`	Finance	Financial advice
`security_expert`	Security	IT security, pen testing

Custom categories can be added via text input.

Assigning to Users¶

Templates are assigned to users via permissions (resource_type: expert_template):

Admin → Edit User → Tab Permissions → Section Expert Template → Select template → +

The user then sees the granted template in the User Portal under My Expert Configuration.

Import / Export¶

All admin templates can be exported as JSON and re-imported.

Export¶

Admin → Expert Templates → Export button

Downloads expert_templates.json. See Import & Export for the full schema.

Import¶

Two ways to import templates:

Option A – Upload file:

Admin → Expert Templates → Import JSON → Select JSON file

Option B – Paste JSON code:

Admin → Expert Templates → Import Code → Paste JSON → Import

Import modes:

Mode	Behavior
`merge`	Templates with the same name are skipped
`replace`	Templates with the same name are overwritten

Result display: X imported, Y skipped

Cluster Impact¶

Action	Effect
Create template	Immediately available in permission management
Edit template	Takes effect on the user's next request using the template
Delete template	All permissions for this template are invalidated
Assign template to user	User's next API request uses the template

Service Toggles¶

Each expert template includes three boolean toggles that control which pipeline components are active for requests using that template:

Toggle	Default	Effect when disabled
`enable_cache`	`true`	Skips ChromaDB L1 semantic cache lookup
`enable_graphrag`	`true`	Skips Neo4j knowledge graph query and ingestion
`enable_web_research`	`true`	Skips SearXNG web search

These toggles are stored as part of the template JSON and propagated through AgentState to each pipeline node. Use cases:

Compliance-sensitive templates: Disable enable_web_research to guarantee no data leaves the local network during query processing.
Speed-optimized templates: Disable enable_graphrag and enable_cache to reduce latency for simple, stateless queries.
Offline templates: Disable all three for pure LLM-only processing with no external dependencies.

Pinned vs Floating Expert Assignment¶

Expert models in a template can be assigned in two modes:

Mode	Endpoint Field	Behavior
Pinned	`model:tag@N04-RTX`	Model always runs on the specified node. Fails if node is unavailable.
Floating	`model:tag` (no `@node`)	Orchestrator searches all configured nodes for the model. Prefers warm nodes.

When to use pinned: Deterministic scheduling, compliance requirements (data must stay on a specific host), or when a model is only installed on one node.

When to use floating: Maximum availability and automatic failover. The orchestrator uses the 3-phase selection strategy (sticky session, warm model, load score) to pick the optimal node at request time. See Floating Node Discovery.

Compliance Badge¶

Each template card in the Admin UI displays a color-coded privacy badge that indicates whether data may leave the local network:

Badge	Color	Condition
Local Only	Green	All expert endpoints resolve to local Ollama nodes
Mixed	Yellow	Some experts use external APIs (e.g., OpenAI, Anthropic)
External	Red	All experts use external APIs

The badge is computed automatically from the template's endpoint configuration. It helps administrators quickly identify which templates are safe for sovereignty-critical workloads.

User-Owned Templates¶

Users with the expert role can create their own templates in the User Portal. These are only visible to the respective user. Admins can view all user templates under /user-content and delete them if needed.