Prompt Engineering for Multilingual Conversations: A Practical Guide
Standard English-only LLMs struggle with code-switched and multilingual queries. Here are the exact techniques we use to make AI agents work reliably across English and local languages.
The Multilingual Problem
"I need to book an appointment for tomorrow at 3pm with Dr. Sharma — but only if she's available in the evening"
This is a perfectly natural query from a WhatsApp user mixing English with informal phrasing or local language fragments. Standard GPT-4o-mini gets the intent but struggles with time expressions, relative dates like "tomorrow", and the conditional nature of the request.
Our Solution: Few-Shot Multilingual Layer
We inject 8–12 multilingual examples into every system prompt for globally deployed agents. These are hand-curated from real customer conversations.
Technique 1: Intent extraction examples
Show the LLM exactly what multilingual queries look like and what structured intent to extract:
User: "when is the doctor available today"
Intent: { type: "availability_check", doctor: null, date: "today" }
User: "can I get a 3pm slot on the 10th"
Intent: { type: "booking_request", time: "15:00", date: "10th" }
Technique 2: Transliteration awareness
LLMs sometimes confuse transliterated words with English words. Explicit disambiguation:
"Mañana" = tomorrow (Spanish)
"Demain" = tomorrow (French)
"OK" / "Got it" = acknowledgement
Results
Before: 58% intent accuracy on multilingual queries.
After 12-example few-shot: 84% intent accuracy.
After regional language routing: 91%.
The remaining 9% are mixed-script queries (non-Latin + Latin scripts) which we now flag for human review.
Writing about AI automation, growing businesses worldwide, and building products that work across languages and markets.
Ready to try it for your business?
7-day free trial. No credit card. Setup in 30 minutes.
Start Free Trial