3) Classification & Routing

Goal

Return stable labels (intent, category, priority) with minimal variance.

Recommended baseline settings

• temperature: 0
• top_p: 1
• max_tokens: small (e.g., 32)
• response_format: json_schema with an enum

Example schema (intent routing):

{
  "name": "Intent",
  "strict": true,
  "schema": {
    "type": "object",
    "additionalProperties": false,
    "properties": {
      "intent": {"type": "string", "enum": ["procurement", "hr", "it", "finance", "other"]},
      "rationale": {"type": "string"}
    },
    "required": ["intent", "rationale"]
  }
}

How to test

• Use a labeled dataset
• Measure accuracy / confusion matrix
• Add a regression test suite so upgrades don’t change routing behavior unexpectedly

This is one of the strongest areas for executive confidence, because structure + low randomness yields stable outputs.