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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Multi-Language Customer Service Agent

ADK Multi-Language Customer Service Agent

Author: Venkata Sudhakar

A single ADK agent can serve customers in multiple Indian languages by detecting the input language and responding appropriately. This is far simpler than maintaining separate agents per language - one instruction set, one deployment, one monitoring dashboard. The Gemini model understands and generates natural text in all major Indian languages, making truly multilingual agents straightforward to build.

ShopMax India deployed a multilingual customer service agent after finding that 42% of its Tier 2 city customers preferred to type in Hindi, Tamil, or Telugu but were forced to use English. The multilingual agent increased message completion rates (customers who got their query resolved) from 61% to 84% for non-English speakers, simply by responding in the language the customer chose to write in.

The below example shows an ADK agent that detects the customer language and responds naturally in the same language.

import vertexai
from google.adk.agents import LlmAgent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService

vertexai.init(project="shopmax-india", location="asia-south1")

multilingual_agent = LlmAgent(
    model="gemini-2.0-flash",
    name="multilingual_cs_agent",
    instruction=(
        "You are a customer service agent for ShopMax India. "
        "IMPORTANT: Always detect the language the customer is writing in "
        "and respond in that same language. "
        "Supported languages: Hindi, Tamil, Telugu, Kannada, Malayalam, "
        "Bengali, Marathi, Gujarati, and English. "
        "If the customer writes in Hindi, respond fully in Hindi. "
        "If they write in Tamil, respond fully in Tamil. "
        "Never switch languages mid-response. "
        "Keep product names, order IDs, and prices in their original form. "
        "Be warm, helpful, and professional in all languages."
    )
)

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

def ask_agent(user_id: str, session_id: str, message: str) -> str:
    result = None
    for event in runner.run(
        user_id=user_id, session_id=session_id,
        new_message={"role": "user", "parts": [{"text": message}]}
    ):
        if event.is_final_response():
            result = event.content.parts[0].text
    return result

# Test with different languages
test_messages = [
    ("CUST-1", "My order ORD-58291 has not arrived yet. What is the status?"),
    ("CUST-2", "मेरा आर्डर ORD-58292 अभी तक नहीं आया। कृपया बताएं।"),
    ("CUST-3", "என் ஆர்டர் ORD-58293 வரவில்லை. என்ன நடக்கிறது?")
]

for uid, msg in test_messages:
    session = session_service.create_session(
        app_name="multilingual_cs", user_id=uid
    )
    response = ask_agent(uid, session.id, msg)
    print(f"Input ({uid}): {msg[:50]}...")
    print(f"Response: {response[:100]}...")
    print()

It gives the following output,

Input (CUST-1): My order ORD-58291 has not arrived yet. What...
Response: I apologise for the delay with your order ORD-58291. Let me check
  the current delivery status for you right away...

Input (CUST-2): मेरा आर्डर ORD-58292 अभी तक नहीं आया। कृपया...
Response: आपके आर्डर ORD-58292 के बारे में असुविधा के लिए हमें खेद है।
  मैं अभी आपके ऑर्डर की स्थिति जांचता हूं...

Input (CUST-3): என் ஆர்டர் ORD-58293 வரவில்லை. என்ன...
Response: உங்கள் ஆர்டர் ORD-58293 தாமதமாவதற்கு மன்னிப்பு கேட்கிறோம்.
  உங்கள் டெலிவரி நிலையை இப்போதே சரிபார்க்கிறேன்...

The below example shows how to add language detection logging to track which languages customers use, informing future localisation investments.

from vertexai.generative_models import GenerativeModel
from collections import Counter

detect_model = GenerativeModel("gemini-2.0-flash")

def detect_language(text: str) -> str:
    response = detect_model.generate_content(
        f"Detect the language of this text. Return ONLY the language name in English "
        f"(e.g. Hindi, Tamil, English). Text: {text[:200]}",
        generation_config={"temperature": 0.0}
    )
    return response.text.strip()

def analyse_language_distribution(messages: list) -> dict:
    language_counts = Counter()
    for msg in messages:
        lang = detect_language(msg)
        language_counts[lang] += 1

total = sum(language_counts.values())
    return {
        lang: {"count": count, "pct": round(count / total * 100, 1)}
        for lang, count in language_counts.most_common()
    }

# Analyse a sample of customer messages
sample_messages = [
    "Where is my order?", "मेरा ऑर्डर कहाँ है?",
    "என் ஆர்டர் எங்கே?", "నా ఆర్డర్ ఏది?",
    "I want to return this product", "मैं वापस करना चाहता हूं"
]

dist = analyse_language_distribution(sample_messages)
for lang, stats in dist.items():
    print(f"{lang}: {stats['count']} messages ({stats['pct']}%)")

It gives the following output,

English: 2 messages (33.3%)
Hindi: 2 messages (33.3%)
Tamil: 1 messages (16.7%)
Telugu: 1 messages (16.7%)

Language distribution analytics give ShopMax India data to drive localisation priorities. When the analytics showed that 28% of support messages came in Hindi and 19% in Tamil, the team prioritised full Hindi and Tamil localisation of the website over other languages. The single multilingual agent eliminates the operational complexity of routing customers to language-specific agents, while the language detection logging turns every customer interaction into a data point for the localisation roadmap.

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