The design problem
A voice AI with a 10,000-row catalogue cannot 'know' the catalogue. It must look things up during the call, in half a second, from a spoken name that may arrive as एल्मेक्स ४०० when the catalogue says Almex 400. And it must never invent a brand that does not exist.
What was built
- Function calling instead of prompt stuffing. The AI calls a lookup for drugs, hospitals and emergency numbers; its prompt holds rules, not data. The voice log shows a zero-character knowledge base by design.
- Brand and generic search over 10,077 brands and 1,225 generics with fuzzy matching, so 'which brands have this molecule' works.
- Devanagari-to-Latin matching. The speech layer often passes names and digits in Devanagari; a transliteration step before the search removed the zero-match cases that had caused invented answers.
- Hospital directory of 349 hospitals with district and province, name-contains and fuzzy search, a browse mode by location, and a ranking that puts real hospitals before clinics and telemedicine listings.
- Emergency directory of 15 blood banks and 11 ambulance services with phone numbers, by area.
- A 'not found' rule. If the lookup returns nothing, the AI says so; it never fills the gap.
What the line does today
Callers ask about a medicine by brand or generic, hear what it is and which alternatives exist, ask for the nearest hospital in their district, or get an ambulance number read out as digit groups. Unresolved questions become tickets. The line went live in June 2026 and handled its first two hundred calls in the first days.
Why this matters for wholesalers
The same lookup pattern is the core of an order desk: a pharmacy reads out items in Nepali, the AI matches each to the catalogue with the authoritative price and scheme, reads the order back and creates it. Pharma and wholesale solution →