Route customer support with Jev

Ask a Choice question for the queue and a separate Score for urgency. A payment complaint may be a billing issue without being urgent, while a production outage may need immediate escalation. One broad label often hides that distinction.

Verify your email and create a key in Console → API keys. Send the JSON below to POST https://jevmodel.app/api/v1/systemone with Authorization: Bearer YOUR_JEVMODEL_KEY and Content-Type: application/json. The same state is evaluated against two separate questions. These queue names and urgency levels are examples: rewrite them to match your actual teams and escalation policy. Keep the key on your server.

{
  "state": "Our payment page is failing for every customer and orders have stopped.",
  "questions": {
    "queue": {
      "type": "choice",
      "instructions": "Which team should handle the reported problem?",
      "criteria": {
        "billing": "Individual charges, refunds or invoice corrections.",
        "technical": "Broken application behavior, API failures or outages.",
        "review": "Unclear, mixed or unsupported requests needing human triage."
      }
    },
    "urgency": {
      "type": "score",
      "instructions": "Rate operational urgency using the ordered levels.",
      "criteria": [
        "Routine request; normal support response is sufficient.",
        "A user is blocked; review during the current support shift.",
        "A widespread outage stops core operations; escalate immediately."
      ]
    }
  }
}

Inspect data.answers.queue and data.answers.urgency in a successful API response. An illustrative queue result could select technical with probabilities technical: 0.86, billing: 0.09, review: 0.05; this is not a recorded run or an 86% correctness guarantee. Score may fall between rubric levels and is not inherently a 0–1 value. Keep urgency separate from queue selection, and choose action thresholds using labeled tickets.

When probabilities are close or the message includes risk signals, send it to a person. Keep the original message attached to the ticket and measure how often the automatic route is later corrected.

Use Console → Batch with CSV, TXT or JSONL inputs to inspect row-level answers and export a CSV. Join those answers to your own reviewer labels outside the site; Batch does not calculate labeled accuracy automatically. In Console → Flows, start from the support template, edit its questions and connect uncertain outcomes to a review outcome. That outcome is a returned value: your own application must create the ticket or notify a person. Only questions visited on the flow path run. A successful flow consumes two daily free runs in total; failed free flows refund both. With fewer than two free runs left, visited questions use actual input tokens, and completed paid nodes remain charged if a later node fails.

The GitHub support-triage directory linked below provides a synthetic billing complaint, a Choice request and an illustrative response. Run the command from the repository root. Add --live after setting JEVMODEL_API_KEY only to send one real request. Inspect data.answers.queue.choice and any returned probabilities; uncertain or unfamiliar categories need human triage. The example does not create tickets or issue refunds.

node quickstart/node.mjs support-triage
# Or: python quickstart/python.py support-triage