Domain Signal

Overview

domain classifies the request topic family. Define domain rules under routing.signals.domains.

The detector uses the classifier configured under global.model_catalog.modules.classifier and its model bindings under global.model_catalog.system.

Key Advantages

• Routes by topic without hard-coding every phrase into keyword lists.
• Keeps domain policy reusable across multiple decisions.
• Supports stable category families that are easy to audit.
• Works well as the first learned signal in a routing graph.

What Problem Does It Solve?

Keyword routing breaks down once prompts are paraphrased or when domain boundaries are broader than a handful of phrases.

domain solves that by mapping topic classification into named routing signals that decisions can compose with complexity, safety, or plugin logic.

When to Use

Use domain when:

• routes are organized around topic families
• lexical matching is too brittle
• the same topic boundary should feed several decisions
• you want a stable learned classifier before adding more specialized signals

Configuration

routing:
  signals:
    domains:
      - name: business
        description: Business and management related queries.
        mmlu_categories: [business]
      - name: law
        description: Legal questions and law-related topics.
        mmlu_categories: [law]
      - name: psychology
        description: Psychology and mental health topics.
        mmlu_categories: [psychology]
      - name: health
        description: Health and medical information queries.
        mmlu_categories: [health]
      - name: other
        description: General fallback traffic.
        mmlu_categories: [other]

Keep domain names stable because decisions reference those names directly.

Only declared domains match. A classifier label matches the domain that lists it in mmlu_categories, or the domain named after it. A label no domain lists counts as other, so the domain that lists other, as above, is the fallback for every topic you didn't declare; without one, such a request matches no domain.

Local and remote classifier selection

With no backend, the domain classifier (Vela 2.0 0.3B by default) runs in the model runtime. The runtime reads the model's architecture from the package, so there is no local selector to set; vllm-sr config migrate removes the earlier variant, use_modernbert and use_mmbert_32k keys.

A remote category classifier uses the shared backend block. Its model is an explicit name from global.model_catalog.external[], and that catalog entry must have model_role: classification. Category currently accepts only the http_classify protocol and the label_distribution.v1 contract so the full label distribution continues to feed domain matching and routing decisions.

global:
  model_catalog:
    external:
      - name: domain-service
        model_role: classification
        llm_endpoint:
          address: domain-classifier.default.svc
          port: 8080
        llm_model_name: domain-intent-v1
    modules:
      classifier:
        domain:
          backend:
            protocol: http_classify
            contract: label_distribution.v1
            model: domain-service
            deadline_ms: 5000

Dependencies and Limitations

Domain classification uses the configured classifier module and processes the request text. Declare other as the fallback, and re-evaluate labels and thresholds when the classifier changes. See a complete example: config/fragments/signal/domain/mmlu.yaml.