What does this project do?

It identifies a customer's primary intent from a free-text support message and returns a structured result with confidence scores and probability distributions, so downstream systems can route the conversation appropriately instead of relying on hardcoded keyword rules.

Which intents does it support?

Twelve: order_status, order_cancel, delivery_tracking, delivery_delayed, delivery_address_change, payment_failed, payment_refund, product_information, product_availability, return_request, return_status and unknown. Each has a human-readable description in domain.py.

What exactly does it return?

An IntentResult with the chosen intent, a confidence between 0.0 and 1.0, a requires_clarification boolean set true when confidence is below 0.8, and a probabilities dictionary mapping every intent to its score for fallback logic.

How do I run the classifier?

Set your OpenRouter API key in a .env file as OPENROUTER_API_KEY, then run uv run python main.py; the entry point loads credentials from the environment, classifies a sample message and prints the structured result.

Does it make the routing decision for me?

No. The README is explicit that the tool provides probabilistic intent detection, not hardcoded rules, and that routing decisions remain configurable — you decide what each intent or low-confidence result should trigger downstream.

Can I give it conversation history?

Yes. Pass prior Turn items, labelled user or assistant, as history to disambiguate short or ambiguous follow-ups; only the latest message is classified, and the history is used purely as context.

How is the code organized?

domain.py holds the pydantic models and the INTENTS mapping, classifier.py wraps the TypeSafe SDK using Choice questions with criteria and the 0.8 threshold, and main.py is the entry point that demonstrates one classification through TypeSafeClient.system_one().

Jev directory

CommunityJev Repos & SDKs

Intent Routing with TypeSafe's JEV Model

Python intent classification for customer support: Jev picks from twelve intents and returns a confidence score plus a per-intent probability distribution, flagging anything below 0.8 confidence as needing clarification while routing stays configurable downstream.

See highlights Open github.com

Stars: 1
Language: Python
Reading time: 2 min
Published by: Community
Author: viniciusfinger
Last verified: Oct 2026

viniciusfinger/jev-intent-classification

User repository on GitHub

View on GitHub

JEV intent classification using Python

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import logicrw intent-classification customer-support

Works forJev for Classification Jev for Routing

Used inJev for Customer Support Jev for E-commerce & Retail

Languages: Python
Platforms: OpenRouter
Capabilities: Open sourceRequires API key

Highlights

Twelve intents from order_status, delivery_tracking and payment_refund to return_request, product_availability and unknown.
Returns intent, confidence 0.0-1.0, a requires_clarification flag below 0.8 confidence, and a probability per intent for fallback logic.
Uses Choice questions with criteria through the typesafe-sdk and a single-step classification call.
Multi-turn context: prior turns pass as history to disambiguate short follow-ups, and only the latest message is classified.
The architecture runs main.py to IntentClassifier.classify to TypeSafeClient.system_one, with the model reached via the OpenRouter API.

Quickstart

python

history = [
    Turn(role="user", content="My package was supposed to arrive yesterday."),
    Turn(role="assistant", content="Sorry about that! How can I help?"),
]
classifier.classify("I just want my money back", history=history)

Watch out

Copy reviewed against the repo README on 2026-10-15; the README is truncated in the digest inside its .env example and publishes no accuracy benchmarks. The model path is named as OpenRouter in the architecture and topics, and the repo provides no routing rules of its own — it returns probabilities for downstream systems.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

The Open Reimplementations: Local Models That Speak System One

Sep 22, 2026

Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.

Jev Repos & SDKs#open-models#research#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai