Complexity Signal

Overview

complexity rates how difficult a request is to answer well: easy, medium or hard. It asks the complexity preset, "How difficult is this request to answer well?", over easy (straightforward retrieval, rewriting or formatting), medium (several steps of reasoning, or a moderate amount of code or analysis) and hard (advanced multi-step reasoning or specialized expertise).

Key Advantages

• Sends hard requests to a reasoning model and easy ones to a fast one.
• Needs no example banks and no thresholds to fit.
• One question per request answers every complexity rule.

What Problem Does It Solve?

Routing every request to the strongest model wastes it on easy work; routing every request to the cheapest fails the hard ones. complexity tells them apart before the model is chosen.

When to Use

Use complexity when a decision should pick a stronger model, or turn on reasoning, for hard requests only. Combine it with domains for "hard math" or "easy code" routes. For a finer scale, write a systemone score question of your own.

Configuration

recipes:
  - name: default
    signals:
      domains:
        - name: math
      complexity:
        - name: difficulty
        - name: needs_reasoning
          description: Multi-step reasoning or synthesis-heavy prompts.
          labels: [hard]
    decisions:
      - name: hard-math
        priority: 200
        rules:
          operator: AND
          conditions:
            - {type: domain, name: math}
            - {type: complexity, name: difficulty, label: hard}
        modelRefs:
          - model: reasoner
      - name: easy-anything
        priority: 100
        rules:
          operator: AND
          conditions:
            - {type: complexity, name: difficulty, label: easy}
        modelRefs:
          - model: fast-model

Field	Meaning
name	The rule's name, which {type: complexity, name: ...} conditions use
labels	The levels the rule matches, among easy, medium and hard; a rule without labels matches nothing itself, and its conditions name a level with label
deployment, timeout_ms	The deployment that answers, and the answer's deadline

A rule matches when the level the model chooses is one of its labels; a condition with a label matches when the model chose that level. A condition can add a predicate on the level's probability.

The model that answers

The decision model answers every complexity rule, unless the complexity binding or the rule's deployment names another; see Choose a deployment. No model has a trained form of the question yet, so every model answers it zero-shot: check it on your own traffic before it gates a costly route. It reads the current user turn, truncated to the model's input.

Configurations written for earlier releases scored example banks against thresholds (hard, easy, threshold, composer, ...) and named a level as difficulty:hard. vllm-sr config migrate rewrites both: the rule keeps its name, and conditions name the level with label. See a complete example: config/fragments/signal/complexity/escalation.yaml.