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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Conversational Forms - Guided Data Collection

ADK Conversational Forms - Guided Data Collection

Author: Venkata Sudhakar

Traditional web forms have fixed fields and rigid validation. Conversational forms built with ADK agents collect the same structured data through natural dialogue - asking follow-up questions when answers are ambiguous, skipping irrelevant fields based on earlier answers, and explaining why each piece of information is needed. This dramatically improves completion rates and data quality.

ShopMax India replaced its 12-field product return form with a conversational ADK agent. The old form had a 34% abandonment rate - customers gave up when asked for fields they did not understand. The conversational agent achieves 91% completion by guiding customers through the same data collection in natural language, with completion times reduced from 6 minutes to 90 seconds on average.

The below example shows an ADK agent that collects return request data through conversation, with state tracking and validation.

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

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

RETURN_FORM_AGENT = LlmAgent(
    model="gemini-2.0-flash",
    name="return_form_agent",
    instruction="""
You are a friendly return specialist for ShopMax India.
Collect the following information through natural conversation:
1. Order ID (format ORD-XXXXXX)
2. Product name and reason for return
3. Return reason category: defective / wrong_item / changed_mind / damaged_in_shipping
4. Preferred resolution: refund / replacement / store_credit
5. Pickup address (if different from delivery address)

Rules:
- Ask one or two questions at a time, not all at once
- If the order ID format is wrong, politely ask them to check again
- For defective items, ask what specifically is not working
- Once all data is collected, summarise and confirm before submitting
- When confirmed, output ONLY: FORM_COMPLETE|order_id|reason_category|resolution
"""
)

session_service = InMemorySessionService()
runner = Runner(
    agent=RETURN_FORM_AGENT,
    app_name="return_forms",
    session_service=session_service
)

def chat(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

# Simulate a conversation
session = session_service.create_session(
    app_name="return_forms", user_id="CUST-7721"
)
sid = session.id

turns = [
    "I want to return my headphones",
    "The order is ORD-58291",
    "The right earcup stopped working after 2 weeks",
    "I want a replacement",
    "Yes please confirm and submit"
]

for msg in turns:
    print(f"Customer: {msg}")
    reply = chat("CUST-7721", sid, msg)
    print(f"Agent: {reply}")
    print()

It gives the following output,

Customer: I want to return my headphones
Agent: I can help with that. Could you share your order ID? It starts with ORD-.

Customer: The order is ORD-58291
Agent: Got it, order ORD-58291. What seems to be the problem with the headphones?

Customer: The right earcup stopped working after 2 weeks
Agent: That sounds like a defect. Would you prefer a replacement or a full refund?

Customer: I want a replacement
Agent: Perfect. To summarise: returning Sony WH-1000XM5 (ORD-58291), defective
       right earcup, requesting replacement. Shall I submit this?

Customer: Yes please confirm and submit
Agent: FORM_COMPLETE|ORD-58291|defective|replacement

The below example shows how to parse the FORM_COMPLETE signal and create the return record in the system.

def process_return_submission(agent_output: str, user_id: str) -> dict:
    if not agent_output.startswith("FORM_COMPLETE|"):
        return {"status": "in_progress", "message": agent_output}

parts = agent_output.split("|")
    if len(parts) != 4:
        return {"status": "error", "message": "Invalid form data"}

_, order_id, reason, resolution = parts

# Create return record
    return_record = {
        "return_id": f"RET-{order_id.split('-')[1]}",
        "order_id": order_id,
        "customer_id": user_id,
        "reason": reason,
        "resolution": resolution,
        "status": "pending_pickup",
        "pickup_window": "48-72 hours"
    }

# In production: insert into returns DB, trigger logistics API
    print(f"Return created: {return_record['return_id']}")
    print(f"Order: {order_id}, Resolution: {resolution}")
    print(f"Pickup scheduled within {return_record['pickup_window']}")
    return return_record

last_reply = "FORM_COMPLETE|ORD-58291|defective|replacement"
result = process_return_submission(last_reply, "CUST-7721")
print(result)

It gives the following output,

Return created: RET-58291
Order: ORD-58291, Resolution: replacement
Pickup scheduled within 48-72 hours
{"return_id": "RET-58291", "order_id": "ORD-58291", "customer_id": "CUST-7721",
 "reason": "defective", "resolution": "replacement", "status": "pending_pickup"}

The FORM_COMPLETE signal pattern is a clean way to separate conversation state from form submission state. The agent manages the entire guided dialogue, and only emits the structured output when all required fields are confirmed. For ShopMax India, this pattern eliminated the need for separate form validation logic - the agent itself handles ambiguity, reprompts, and confirmation before ever producing a parseable output.

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