Skip to main content

Lab 4 Challenge: Customer Support Analyzer

Goal​

Your task is to build a "Support Analyzer" agent. Instead of returning plain text, the agent must analyze the user's support ticket and return a JSON object containing the categorization, sentiment, and a brief summary of the issue.

Lab Tasks​

  1. Open your terminal, ensure you are in your adk-training directory, and run the following command to scaffold a new agent: 
    adk create support_analyzer
  2. Navigate into the newly created support_analyzer directory.
  3. Define a Pydantic Schema: In agent.py, create a class named SupportAnalysis that defines three fields:
    • category: A string (e.g., "billing", "technical", "general").
    • sentiment: A string (e.g., "positive", "negative", "neutral").
    • summary: A string containing a 1-sentence summary of the user's issue.
  4. Update the Agent Configuration:
    • Set the output_schema to your SupportAnalysis class.
    • Set the output_key to "last_ticket_analysis".
  5. Refine Instructions: Update the agent's instruction to ensure it understands its job is to extract these three specific pieces of information from the user's input.
  6. Run and Verify: Run the agent using adk web support_analyzer (from the parent directory). In the Dev UI, test it with a phrase like "My screen is completely broken and I'm very angry about it!" and verify the response is a valid JSON object.
  7. Inspect State: After a few interactions, check the "Session State" in the Dev UI to confirm that "last_ticket_analysis" is being populated correctly.

Python Approach (Primary)​

Modify agent.py to include the Pydantic model and updated LlmAgent parameters.

from pydantic import BaseModel
from google.adk.agents import LlmAgent

# TODO: Define your Pydantic schema
class SupportAnalysis(BaseModel):
    # TODO: Add fields (category, sentiment, summary)
    pass

root_agent = LlmAgent(
    name="support_analyzer_agent",
    model="gemini-2.5-flash",
    instruction="...", # TODO: Update instructions for categorization, sentiment, and summary
    output_schema=...,  # TODO: Link the schema
    output_key=...      # TODO: Set the output key
)

Self-Reflection Questions​

  • Why is it better to use output_schema instead of just asking the LLM to "respond in JSON" in the text instructions?
  • What happened to the agent's ability to use tools once you enabled output_schema? Why does the ADK enforce this restriction?
  • Look at the Dev UI's "Session State". How could another agent in a future multi-agent system use the data stored in the "last_ticket_analysis" key?

🕵️ Hidden Solution 🕵️​

Looking for the solution? Here's a hint (Base64 decode me): L2RvYy1hZGstdHJhaW5pbmcvbW9kdWxlMDQtbGxtYWdlbnQtZGVlcC1kaXZlL2xhYi1zb2x1dGlvbg==

The direct link is: Lab Solution