Intent Router Node

The Intent Router Node classifies the user's input with the Jev decision model (TypeSafe System One, via OpenRouter) and routes execution to one of its class handles — no LLM prompt, no text parsing. One typed decision call, typically 70-500ms.

How it works

1. The node sends the input text plus your class definitions to Jev as a single Choice question.
2. Jev returns the chosen class with calibrated confidence and a probability for every class.
3. The flow routes to the route:<class_id> edge whose class was chosen. If confidence is below your threshold, it routes to the route:low_confidence handle instead.

Parameters

• question (string): The routing question Jev answers (e.g. "Which category does this user request belong to?").
• classes (array): One entry per route. Each class has:
  • id (string): The handle identifier — wire the edge from route:<id>.
  • description (string): What belongs in this class. Descriptions, not names, drive zero-shot quality.
• confidence_threshold (number, 0-1, default 0.7): Below this, the low-confidence handle is used.
• input_key (string, default "user_message"): Which state field to classify.

Output Handles

• route:<class_id> — one per class.
• route:low_confidence — fallback when the decision is not confident enough. Wire this to an LLM-based router or a human review path.

State

Writes intent = { id, confidence, probabilities } — readable by a Condition Node via dotted paths like intent.confidence.

Use Cases

1. Intent routing: UserChatInput → IntentRouter → {billing_flow, support_flow, sales_flow} with a low-confidence fallback to a general agent.
2. Language/team triage: pick the department that should own a ticket.
3. Topic gating: separate FAQ-style questions from deep research requests.

Tips

• Always add an "other"/catch-all class so Jev can say none of the classes fit.
• Write descriptions that separate the classes from each other ("Invoices, payments, refunds" vs "Delivery status, delays").
• If the decision service is unavailable, the node routes to low_confidence — the flow never dead-ends.