Intent Engine

Classifies what a caller or chat user is trying to do — a bank-dispute flow, a connectivity complaint, a yes/no confirmation — without a round trip through an LLM. Built for IVR and voice-bot turns where you need a routing decision in single-digit milliseconds and a cost per call far below a chat completion, reserving the LLM for free-form conversation rather than classification.

How it works

Every call to /v1/understand runs the same pipeline against your text:

classifier_used tells you which stage actually resolved the turn — deterministic, local_model, context (the yes/no shortcut above), llm_fallback, or none when nothing resolved it. It never names a specific third-party model, the same abstraction every other model type in this gateway holds to.

Request body (JSON)

Example

curl https://api.inferexai.in/v1/understand \
  -H "Authorization: Bearer sk-live-your-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "inferex-intent-demo",
    "session_id": "a3f9c2e1-...",
    "text": "ATM se paisa nahi nikla par account se kat gaya",
    "language": "hi-IN"
  }'

{
  "model": "inferex-intent-demo",
  "session_id": "a3f9c2e1-...",
  "request_id": "req_abc123",
  "classification": "KNOWN",
  "intent": { "id": "BANKING.ATM.CASH_WITHDRAWAL.CASH_DEBITED_NOT_RECEIVED", "confidence": 0.86, "band": "HIGH" },
  "candidate_intents": [
    { "id": "BANKING.ATM.CASH_WITHDRAWAL.CASH_DEBITED_NOT_RECEIVED", "score": 0.86, "source": "rules" }
  ],
  "entities": {},
  "dialogue_act": "STATEMENT",
  "decision": {
    "action": "CREATE_DISPUTE",
    "next_state": "AWAITING_ACCOUNT_NUMBER",
    "response_id": "banking.atm.cash_debited_not_received.ask_account"
  },
  "response": { "response_id": "banking.atm.cash_debited_not_received.ask_account", "language": "hi-IN", "channel": "text" },
  "classifier_used": "deterministic",
  "latency_ms": 4
}

Classification values

intent is non-null if and only if classification is KNOWN — there is no "low-confidence guess" to accidentally read as a match.

Entities

entities is a flat object keyed by type — currently amount, phone, email, and date — populated when found in the turn, omitted otherwise:

{
  "...": "...",
  "entities": { "amount": 5000, "phone": "9876543210" }
}

Multi-turn / confirmation

When a decision.next_state indicates the pack is waiting on a yes/no, send the next turn with the same session_id — a bare confirmation is resolved from context, not re-classified, and decision.action/response_id on the confirmed path can differ from the initial ask:

{
  "model": "inferex-intent-demo",
  "session_id": "a3f9c2e1-...",
  "text": "haan",
  "language": "hi-IN"
}

Response is a reference, not text

response.response_id identifies a reply template — it is not the reply text itself. Resolve it (in response.language, over response.channel: text, voice, …) against your own copy store, or ask us to ship your pack with its response copy included so you can request the rendered string in a future version. This keeps wording changes (tone, script updates, a copy fix) out of the classification contract entirely.

Cost

Billed per call to /v1/understand — one charge regardless of which cascade stage resolved the turn (the deterministic stage resolves the large majority of turns in-domain, which is what keeps the per-call cost well below an LLM chat completion). The live rate for your account is on GET /v1/models / the Billing page, same as every other model type — there is no separate pricing surface for intent.

Error codes