Classify requests

Classifiers give every request a label: its subject, whether it needs fact checking, how the user reacts to the last answer, or whether they ask for an image. Routes match on those labels.

Signal	Model	Labels
domain	Vela 1.0 Domain	14 domains: biology, business, chemistry, computer science, economics, engineering, health, history, law, math, other, philosophy, physics, psychology
fact_check	Vela 1.0 FactCheck	FACT_CHECK_NEEDED, NO_FACT_CHECK_NEEDED
user_feedback	Vela 1.0 Feedback	satisfied, need clarification, wrong answer, want different, no feedback
modality	Vela 1.0 Modality	AR (text), DIFFUSION (image), BOTH
classifier	your own model	your labels

Turn it on

Add the signal and use it in a route. The router runs the right model for you on the CPU; nothing else is needed:

routing:
  signals:
    domains:
      - name: math
        description: Mathematics and quantitative reasoning.
        mmlu_categories: [math]
      - name: computer science
        description: Programming and computer science.
        mmlu_categories: [computer science]
  decisions:
    - name: math-route
      priority: 100
      rules:
        operator: AND
        conditions:
          - type: domain
            name: math
      modelRefs:
        - model: math-model

Choose where it runs

To pick the device, the input limit or another model, describe a deployment and bind the feature to it. The binding names are domain_classifier, fact_check_classifier, feedback_detector and modality_detector; all of them read label probabilities (label_distribution.v1):

global:
  model_catalog:
    deployments:
      vela-domain:
        provider: model_runtime
        artifact: vllm-sr/Vela-1.0-Encoder-307M-Domain
        device: cpu
        input:
          max_tokens: 2048
          overflow: truncate
    bindings:
      domain_classifier:
        deployment: vela-domain
        contract: label_distribution.v1

overflow: truncate classifies the first 2,048 tokens of a long request and reports that it did; reject (the default) leaves the signal unknown for input over the limit instead.

Check it

Ask the model directly. Start it on its own, or use any runtime that serves it:

vllm-sr serve vllm-sr/Vela-1.0-Encoder-307M-Domain --device cpu --port 8100
curl -s localhost:8100/v1/classify -H 'content-type: application/json' \
  -d '{"input": ["What is the derivative of x squared?", "Fix this segfault in my C code."]}'

Each result has the most likely label and the probabilities of all labels in the order of labels. Through the router, send a request with the x-vsr-debug: true header: x-vsr-matched-domains lists the domains that matched and x-vsr-selected-decision names the route.

Use your own classifier

A Hugging Face ModernBERT or mmBERT sequence classifier works as a classifier signal with your own labels. Pin its revision, bind it to the rule and route on its labels:

routing:
  signals:
    classifiers:
      - name: ticket_topic
        type: local
        labels: [billing, shipping, other]
  decisions:
    - name: billing-desk
      priority: 100
      rules:
        operator: AND
        conditions:
          - type: classifier
            name: ticket_topic
            label: billing
            predicate:
              gte: 0.5
      modelRefs:
        - model: support-model
  model_bindings:
    classifier.ticket_topic:
      deployment: ticket-topics
      contract: label_distribution.v1
global:
  model_catalog:
    deployments:
      ticket-topics:
        provider: model_runtime
        artifact: your-org/ticket-topic-classifier
        revision: 0123456789abcdef0123456789abcdef01234567
        device: cpu
        input:
          max_tokens: 512
          overflow: reject

List labels in the model's own label order (its id2label). The router reads the model's labels from the runtime and refuses to start if they do not match the rule. A model with independent labels and a published operating point uses label_scores.v1 instead; see classifier signals.