Lab 7: Building a Visual Product Catalog Analyzer

Goal

In this lab, you will build a multimodal agent that can analyze a product image and generate a marketing description. This will teach you how to handle image inputs and perform programmatic execution with a vision-capable agent.

The Architecture

You will use the App and Runner pattern you learned in Module 6 to build a Python script that:

1. Loads local image files as types.Part objects.
2. Constructs a multimodal prompt containing both text and image data.
3. Executes the agent using run_async.

Step 1: Create and Prepare the Project

Note: unlike other modules, visual_catalog must live directly inside this module's own folder (module07-multimodal-and-images/) rather than in your shared adk-training project -- see point 3 below for why. If this is the first agent you're creating from inside this folder, set up a local environment here first: uv init --python 3.10 && uv add "google-adk>=2.1.0" python-dotenv.

1. Create the agent project:

   uv run adk create visual_catalog

2. Navigate into the new directory:

   cd visual_catalog

3. Sample images: two sample product photos, headphones.jpg and laptop.jpg, are already provided in this module's folder (module07-multimodal-and-images/), one level up from visual_catalog/. That's why you must create visual_catalog directly inside this module's folder (as done in step 1 above) — the script below loads them with a relative ../ path. Feel free to swap in your own images later, as long as the filenames match.
4. Set up your .env file. Vision models require a Vertex AI configuration. Ensure your .env file looks like this:

   GOOGLE_GENAI_USE_VERTEXAI=1
   GOOGLE_CLOUD_PROJECT="your-project-id"
   GOOGLE_CLOUD_LOCATION="us-central1"

Step 2: Implement the Multimodal Script

Exercise: Create a file named main.py in your visual_catalog directory. Your task is to complete the core logic inside the analyze_product method by filling in the # TODO sections.

Note on Sessions: When using run_async programmatically, you must explicitly create a session in the runner's session service before sending a message, just as you did with curl in Module 5.

import asyncio
import logging
import os
from google.adk import Agent
from google.adk.apps import App
from google.adk.runners import InMemoryRunner
from google.genai import types
from dotenv import load_dotenv

# Optional: Suppress noisy ADK/httpx logs
logging.getLogger("google.adk").setLevel(logging.WARNING)

# Helper function to load an image from a local file path
def load_image_from_file(path: str) -> types.Part:
    """Load image from file and return a types.Part object."""
    with open(path, 'rb') as f:
        image_bytes = f.read()
    
    mime_type = 'image/png' if path.lower().endswith('.png') else 'image/jpeg'

    return types.Part(
        inline_data=types.Blob(data=image_bytes, mime_type=mime_type)
    )

class VisualCatalogApp:
    def __init__(self):
        # 1. Define the Intelligence
        self.agent = Agent(
            model='gemini-3.5-flash',
            name='catalog_writer',
            instruction="You are a professional product catalog writer. Analyze the image and write a compelling description."
        )
        # 2. Build the Infrastructure
        self.app = App(name="visual_catalog", root_agent=self.agent)
        self.runner = InMemoryRunner(app=self.app)

    async def analyze_product(self, product_id: str, image_path: str):
        print(f"\n--- Analyzing Product: {product_id} ---")
        user_id = "catalog_admin"
        session_id = f"sess_{product_id}"

        # TODO: 1. Create the session explicitly in the session service
        # Hint: Use self.runner.session_service.create_session(...)
        ...

        # TODO: 2. Load the image using the helper function
        image_part = ...

        # TODO: 3. Create the multimodal message (types.Content object)
        msg = types.Content(
            role="user",
            parts=[
                types.Part(text=f"Analyze this image for product {product_id}."),
                image_part
            ]
        )

        # TODO: 4. Run the agent using run_async
        print("📸 Sending image to Gemini...")
        async for event in self.runner.run_async(
            user_id=user_id,
            session_id=session_id,
            new_message=msg
        ):
            if event.is_final_response():
                print(f"✅ Description:\n{event.content.parts[0].text}\n")

async def main():
    load_dotenv()
    catalog = VisualCatalogApp()

    # Images are expected to be in the parent directory for this lab
    products = [
        ('HEADPHONES-01', '../headphones.jpg'),
        ('LAPTOP-02', '../laptop.jpg'),
    ]

    for product_id, path in products:
        if os.path.exists(path):
            await catalog.analyze_product(product_id, path)
            await asyncio.sleep(1)
        else:
            print(f"⚠️ Image not found: {path}")

if __name__ == '__main__':
    asyncio.run(main())

Cleanup: Unlike other modules, this lab's project was created directly inside module07-multimodal-and-images/ (the actual course repo folder) rather than in your separate adk-training scratch directory. Once you're done, delete visual_catalog/, .venv/, pyproject.toml, and uv.lock from inside this module's folder so they don't accidentally get committed to the repo.

Lab Summary

You have successfully built a multimodal agent! You have learned:

• How to package image bytes into a types.Part object.
• The importance of explicitly creating a Session when using run_async programmatically.
• How to construct a structured types.Content object for complex inputs.

This completes Part 1: Foundations. You now know how to build, configure, run, and feed multimodal input to an agent. In Part 2, you'll start extending your agent's capabilities with tools.

Self-Reflection Questions

• Why did we have to call create_session manually this time, whereas in Module 6's run_debug we didn't?
• How does the InMemoryRunner simplify the setup of the SessionService compared to a base Runner?
• If you wanted to analyze a PDF document instead of an image, which mime_type would you use in the types.Blob?

---

🔐 Reference Solution Vault

🔐

Module 07 Reference Solution Vault

Active Comprehension Gate • Zero Guessable URLs • Guided Verification

Step 1/3: Question 1 of 3

How are multimodal assets (images, PDFs, audio) passed to Gemini models in ADK?