@ankr/sahayak-core

Sahayak (Hindi: the helper) — a universal assistant core: one intent layer for every channel (email, WhatsApp, Telegram, UI, voice) and every product.

import { match, resolveRoute } from '@ankr/sahayak-core';

const r = match('please approve the DA payment of USD 46,250 ref DA-FUJ-001', { domain: 'maritime' });
// → { intent: 'APPROVE_REQUEST', act: 'APPROVE', confidence: 0.97, band: 'act',
//     entities: [amount, reference], lang: 'en', route: null }

Why it's different

• Rules before models — 130+ curated patterns (en + hi/Hinglish) match first; an LLM fallback fires only below the suggest band, and its confidence is capped at 0.79 — a model guess can suggest, never auto-act.
• Capabilities are looked up, never claimed — resolveRoute() resolves against a service's declared capability surface (ANKR Forja STATE); an unroutable intent returns route: null, never an invented endpoint.
• Writes are proposals — destructive acts (APPROVE/CANCEL) need confidence ≥ 0.95, and execution always belongs to the consumer's human-gated engine.
• Clarify is a first-class outcome — below 0.60 you get clarify_options (top-3), not a wrong guess.
• The golden set is the contract — every release gates on eval accuracy ≥85%; every real-world miss enters the golden set before any pattern fix.

API

match(text, {domain?}) → MatchResult · resolveRoute(result, surface, overrides?, onUnroutable?) · llmFallback(text, lang, domain?) · extractEntities(text) · detectLanguage(text)

License

AGPL-3.0-only · ANKR Labs (Powerpbox IT Solutions Pvt Ltd)