Skip to main content

​
Prompts for Non-English

Creating effective agents in languages other than English requires specific prompting strategies.

​
1. Language Instruction

Even if you are prompting in English, clearly state the output language.
  • “Prompt Example: ‘You are an assistant. You must speak only in Odia. Do not use English words unless they are technical terms with no Odia equivalent.‘“

​
2. Transliteration vs. Native Script

Most LLMs (like GPT-4) understand both Native Script (e.g., नमस्ते ) and Transliterated text (e.g., “Namaste”).
  • Tip: Provide examples in the Native Script within your prompt to ensure the LLM understands the correct formal/informal tone of that specific culture.

​
3. Cultural Nuance

Languages carry cultural weight.
  • If prompting for a Japanese agent, instruct it on the level of politeness (Keigo).
  • If for a Hindi agent, specify if it should be formal (“Aap”) or semi-formal (“Tum”).

​
4. Testing for ASR Accuracy

Non-English speech recognition can be sensitive to “Hinglish” or “Odia-English” mixing.
  • Instruct your agent to be “Robust to code-switching” if you expect your users to mix languages naturally.
For the best results in non-English (especially Indian) languages, use Sarvam AI for both ASR and TTS.