	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > Multilingual Agent - Hindi and English

Multilingual Agent - Hindi and English

Author: Venkata Sudhakar

India is a multilingual market. ShopMax India serves customers across Hindi-speaking states like UP, Bihar, MP, and Rajasthan where many prefer communicating in Hindi, as well as metro customers who prefer English. A Multilingual Agent detects the language of each message and responds accordingly - no manual routing needed. Gemini natively understands Hindi written in Devanagari script as well as Hinglish (Hindi written in Roman script), making it ideal for Indian language support.

This tutorial builds a Multilingual Customer Service Agent using Google ADK and Gemini 2.0 Flash. The agent detects the input language, fetches the appropriate response in that language, and maintains language consistency throughout the conversation.

The below example shows the Multilingual Agent handling a product enquiry first in Hindi then switching to English.

from google.adk.agents import Agent
from google.adk.tools import FunctionTool
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.genai.types import Content, Part

# Product catalogue (bilingual)
PRODUCTS = {
    "laptop": {
        "name_en": "Dell Latitude 15 Laptop",
        "name_hi": "Dell Latitude 15 Laptop",
        "price": 68000,
        "warranty_en": "3 years on-site warranty",
        "warranty_hi": "3 saal ki on-site warranty",
        "stock": "available"
    },
    "printer": {
        "name_en": "HP LaserJet Pro Printer",
        "name_hi": "HP LaserJet Pro Printer",
        "price": 22000,
        "warranty_en": "1 year carry-in warranty",
        "warranty_hi": "1 saal ki carry-in warranty",
        "stock": "available"
    }
}

def detect_language(text: str) -> dict:
    """Detect if the input is Hindi (Devanagari), Hinglish, or English."""
    # Check for Devanagari Unicode range
    devanagari_chars = sum(1 for c in text if "\u0900" <= c <= "\u097F")
    total_chars = len(text.replace(" ", ""))
    if devanagari_chars > 0:
        lang = "hindi_devanagari"
    else:
        # Hinglish detection - common Hindi words in Roman script
        hinglish_words = ["kya", "hai", "mujhe", "chahiye", "kitna", "price",
                          "batao", "bata", "karo", "nahi", "hoga", "kaisa"]
        text_lower = text.lower()
        hinglish_count = sum(1 for w in hinglish_words if w in text_lower)
        lang = "hinglish" if hinglish_count >= 2 else "english"
    return {"detected_language": lang, "respond_in": lang}

def get_product_info(product_key: str, language: str) -> dict:
    """Fetch product info in the requested language."""
    product = PRODUCTS.get(product_key.lower())
    if not product:
        if language in ["hindi_devanagari", "hinglish"]:
            return {"error": "Yeh product hamare paas nahi hai."}
        return {"error": "This product is not available in our catalogue."}
    if language in ["hindi_devanagari", "hinglish"]:
        return {
            "product": product["name_hi"],
            "price": "Rs " + str(product["price"]),
            "warranty": product["warranty_hi"],
            "availability": "Uplabdh hai" if product["stock"] == "available" else "Stock mein nahi"
        }
    return {
        "product": product["name_en"],
        "price": "Rs " + str(product["price"]),
        "warranty": product["warranty_en"],
        "availability": "In stock" if product["stock"] == "available" else "Out of stock"
    }

lang_tool = FunctionTool(func=detect_language)
product_tool = FunctionTool(func=get_product_info)

multilingual_agent = Agent(
    model="gemini-2.0-flash",
    name="multilingual_agent",
    instruction="""
        You are a bilingual customer service agent for ShopMax India.
        For every message:
        1. Detect the language of the customer input
        2. Fetch product info in that language if a product is mentioned
        3. Always respond in the SAME language as the customer
           - If Hindi/Devanagari: respond in Hindi
           - If Hinglish: respond in friendly Hinglish
           - If English: respond in professional English
        4. Be warm, helpful, and match the customer tone
        Never mix languages unless the customer does first.
    """,
    tools=[lang_tool, product_tool]
)

session_service = InMemorySessionService()
runner = Runner(agent=multilingual_agent, app_name="multilingual", session_service=session_service)

# Test Hindi query
session = session_service.create_session(app_name="multilingual", user_id="customer1")
queries = [
    "Laptop ka price kitna hai aur warranty kya milegi?",
    "What is the price of the HP printer?"
]
for query in queries:
    print("Customer:", query)
    message = Content(role="user", parts=[Part(text=query)])
    for event in runner.run(user_id="customer1", session_id=session.id, new_message=message):
        if event.is_final_response():
            print("Agent:", event.content.parts[0].text)
    print()

It gives the following output,

Customer: Laptop ka price kitna hai aur warranty kya milegi?
Agent: Namaste! Dell Latitude 15 Laptop ki jaankari:
  Daam    : Rs 68,000
  Warranty: 3 saal ki on-site warranty
  Stock   : Uplabdh hai
  Kya aap order karna chahte hain ya koi aur sawaal hai?

Customer: What is the price of the HP printer?
Agent: Thank you for your enquiry!
  HP LaserJet Pro Printer details:
  Price       : Rs 22,000
  Warranty    : 1 year carry-in warranty
  Availability: In stock
  Would you like to place an order or need more information?

The Multilingual Agent can be extended to support Tamil, Telugu, Kannada, Bengali, and Marathi - Gemini understands all major Indian languages. Deploy it as a WhatsApp Business bot to reach customers in their native language at scale. For voice interactions, combine it with the Gemini Live API to handle Hindi and regional language calls without human agents.

Send your comments, suggestions or queries regarding this site to [email protected].