Prompt Engineering for Hinglish: A Practical Guide
Standard English LLMs hallucinate on Hinglish queries. Here are the exact techniques we use to make AI agents work reliably in India's most common conversational register.
The Hinglish Problem
"Kal ka appointment 3 baje ke liye book karna hai, Doctor Sharma ke saath — but only if she is available in the evening"
This is a perfectly natural query from an Indian WhatsApp user. Standard GPT-4o-mini gets the intent but struggles with "3 baje" (3 o'clock), "kal" (tomorrow), and the conditional nature of the request.
Our Solution: Few-Shot Hinglish Layer
We inject 8–12 Hinglish examples into every system prompt for India-deployed agents. These are hand-curated from real clinic conversations.
Technique 1: Intent extraction examples
Show the LLM exactly what Hinglish queries look like and what structured intent to extract:
User: "doctor ka time kab hai, aaj"
Intent: { type: "availability_check", doctor: null, date: "today" }
User: "3 baje slot milega kya 10 tarikh ko"
Intent: { type: "booking_request", time: "15:00", date: "10th" }
Technique 2: Transliteration awareness
LLMs sometimes confuse Hindi transliterations with English words. Explicit disambiguation:
"Kal" = tomorrow (not the name Kal)
"Abhi" = right now (not a name)
"Theek hai" = OK/understood (acknowledgement)
Results
Before: 58% intent accuracy on Hinglish queries.
After 12-example few-shot: 84% intent accuracy.
After Sarvam AI routing for native Hindi: 91%.
The remaining 9% are mixed-script queries (Devanagari + Roman) which we now flag for human review.
Writing about AI automation, India SMBs, and building products that work for the next billion users.
Ready to try it for your business?
7-day free trial. No credit card. Setup in 30 minutes.
Start Free Trial