n8n-nodes-jev-router

An n8n community node integrating TypeSafe AI's
Jev model — a "System One" decision model, not a chat model. You send it one or more typed
questions (Choice, Score, or Noul/Boolean) about a piece of text or data, and it answers each
with a structured, typed value plus a confidence or probability score, in a single API call.

This package integrates exactly one third-party service — TypeSafe AI's /v1/systemone
endpoint — through one node with two operations: Classify & Route, which calls Jev and
optionally branches the workflow by one of its answers, and Calibration Check, which
validates Jev's confidence scores against your own labeled data. The branching in Classify &
Route is a UX layer built directly on top of that API call (similar to how many integration
nodes expose a success/error output), not a standalone flow-control utility.

What this node does

• Classify & Route — send an item's text/data to Jev along with one or more questions,
  answered in a single API call. Optionally pick one Choice question as the "routing question":
  the node then creates one output branch per possible answer (like the built-in Switch node),
  plus a Needs Review branch that catches anything below that question's confidence
  threshold. Every branch still gets all answers attached to the item, so you can reference
  a Score or Noul answer downstream even if it wasn't the one used for routing.
• Calibration Check — before you trust Jev's confidence scores in production, run this
  mode against your own labeled historical data. It compares Jev's prediction to a ground-truth
  field on each item, buckets the results by confidence range, and reports accuracy per bucket
  — so you can pick a real threshold (e.g. "above 0.85 confidence, Jev was right 96% of the
  time") instead of guessing 0.5 and hoping.

The node is also flagged usableAsTool, so it can be added directly to an n8n AI Agent
node's tool list and called by an agent.

Install

In n8n: Settings → Community Nodes → Install, and enter:

n8n-nodes-jev-router

For local development against a self-hosted n8n instance, set N8N_CUSTOM_EXTENSIONS to this
project's dist/ directory (after npm install && npm run build) so n8n picks it up without
needing the package published first — see
n8n's community node development docs for
details.

Credentials

Create a Jev API credential with:

• API Key — your TypeSafe AI API key, sent as Authorization: Bearer <key>.
• Base URL — defaults to https://api.typesafe.ai. Override this to point at OpenRouter
  or another Jev-compatible gateway.

Example: routing support tickets

A common use case is triaging inbound support tickets without an LLM chat round-trip.

Operation: Classify & Route
State: {{$json.ticketBody}}

Questions:

Routing Question: department, with a confidence threshold of 0.75.

A ready-to-import version of this exact workflow (Manual Trigger → sample ticket → Jev Router
→ one Set node per branch, so you can see what lands where) is in
examples/support-ticket-routing.json. In n8n, go to
Workflows → Import from File and select it — you'll just need to pick your own Jev API
credential on the Jev Router node before running it.

This produces four outputs: billing, sales, technical, and Needs Review. A
ticket Jev is confident is billing-related routes straight to the billing branch; anything
under 0.75 confidence — regardless of which department it guessed — lands in Needs Review
for a human to triage. Every branch also carries urgency_score and refund_requested_noul
on the item, so a downstream node can, say, page someone when urgency_score is high even on
a ticket that got auto-routed to sales.

Example: checking calibration before you trust it

Before wiring the routing threshold above into production, run the same questions in
Calibration Check mode against a set of tickets you've already labeled by hand:

Operation: Calibration Check
State Field Name: ticketBody
Ground Truth Field Name: actual_department
Questions: same department question as above
Question to Compare: department

The output is a single item summarizing accuracy per confidence bucket (0.9–1.0, 0.8–0.9, …),
so you can see exactly where Jev's confidence stops being trustworthy on your data before
picking a threshold — rather than assuming the model's stated confidence is well-calibrated
out of the box.

Notes

• Questions are always batched into a single Jev API call per item (never one call per
  question).
• Transient errors (HTTP 429 rate-limited, 529 overloaded) are retried automatically with
  exponential backoff, up to 4 attempts — no node setting to enable.
• Noul questions have no confidence field in Jev's response (only a probability), so the
  Confidence Threshold field is hidden for them in the UI; threshold on the noul probability
  value itself instead.

License

MIT