@o-lang/intent-classifier

O-Lang intent classifier resolver — topic gating for governed chatbots.

Classifies user messages against a list of allowed topics. Blocks off-topic messages before they reach the LLM. Works for any domain — topics are defined at runtime.

Install

npm install @o-lang/intent-classifier

Actions

intent_classifier — classify and gate

Step 1: Ask intent-classifier "{user_message}" "{allowed_topics}"
Save as intent

Returns:

{
  "intent":     "ask_refund",
  "topic":      "refunds",
  "is_allowed": true,
  "confidence": 0.92
}

Use in a workflow

Workflow "My Chatbot" with session_id, user_message, allowed_topics

Allow resolvers:
  - @o-lang/intent-classifier
  - @o-lang/llm-groq

Step 1: Ask @o-lang/intent-classifier "{user_message}" "{allowed_topics}"
Save as intent

Step 2: Ask @o-lang/llm-groq "Answer this: {user_message}"
Save as reply

Return reply

Works for any domain

# Banking
allowed_topics: "accounts,savings,loans,transfers"

# E-commerce
allowed_topics: "orders,refunds,shipping,products"

# Healthcare
allowed_topics: "appointments,prescriptions,billing"

# HR
allowed_topics: "leave,payroll,benefits,onboarding"

LLM vs keyword fallback

| Condition | Behaviour | |---|---| | GROQ_API_KEY set | Uses llama-3.1-8b-instant for accurate classification | | No API key | Falls back to keyword matching — works offline | | LLM error | Automatically falls back to keyword matching |

Environment variables

GROQ_API_KEY=your_key      # optional — enables LLM classification
INTENT_MODEL=llama-3.1-8b-instant  # optional model override

Running tests

npm test                     # Jest suite — no API key required
node test-local-resolver.js  # Live smoke test

Contact

[email protected]