Building agentic RAG for e-commerce with ADK and Vector Search

Key Concepts

• ADK (Agent Development Kit): A toolkit for building AI agents.
• Vati Vector Search: A vector search technology used for efficient data retrieval.
• Retail Data Set: A collection of data related to retail products and customer interactions.
• AI Agent: An intelligent software program designed to perform tasks autonomously.
• Gemini: A large language model (LLM) that powers the AI agent, providing common sense and contextual understanding.
• Deep Research: An advanced search capability that goes beyond simple keyword matching, utilizing external search engines like Google.
• Generative Recommendation System: A recommendation system that uses AI agents and LLMs to provide more contextually relevant and personalized suggestions.
• Traditional Machine Learning Systems: Recommendation systems based on historical purchase or click data.

Surplus Cons Demo: ADK with Vati Vector Search for Retail

This demonstration showcases the ADK (Agent Development Kit) integrated with Vati Vector Search for a retail data set. The core of the demo is an AI agent named "Shoppers Concierge," designed to assist users in an e-commerce environment with approximately 10 million items.

Interactive Product Search and Deep Research

The Shoppers Concierge agent can handle direct product search queries. For instance, when asked to find a birthday present for a 10-year-old son, the agent presents a list of relevant items. A key feature highlighted is the ability to "refine or change the search or start a deep research."

Deep Research Process: When "deep research" is initiated, the agent leverages external search engines, specifically Google in this demo, to gather more comprehensive information. This process involves:

• Generating Multiple Queries: For a single user request, the agent can generate up to 100 sub-queries to explore various aspects of the search. This is demonstrated by the console output showing numerous generated queries.
• Broadening Search Scope: The agent researches popular items that people are buying for similar occasions (e.g., birthday presents for sons).
• Categorizing Results: The deep research results are presented in categorized lists, such as "stem building kits," "outdoor active play equipment," "creative art supplies," "board games and puzzles," and "books and media."

This deep research capability aims to alleviate the burden on users who would otherwise need to type multiple, varied queries themselves.

"Kazier's Picks" and Image-Based Contextual Search

The demo introduces "Kazier's Picks," which are featured items selected from the deep research results.

Image-to-Product Contextualization: A significant advancement demonstrated is the agent's ability to understand context from user-provided images.

• Example: A user can upload an image of their home office or take a picture of a desk with a laptop and chair.
• Agent's Understanding: The agent analyzes the image to identify relevant items. For instance, if an image of a home office is provided, the agent can suggest home office accessories.
• Product Details: Upon selecting an item, such as an "LED lamp," the agent provides detailed product descriptions, including features like a USB charging port, three-way touch switch, and adjustable neck.

Generative Recommendations vs. Traditional Systems

The demo contrasts the AI agent-powered system with traditional e-commerce recommendation engines.

Traditional Systems:

• Rely on traditional machine learning models.
• Utilize statistics from past purchase history or click history.

Generative Recommendation System (AI Agent Powered):

• Powered by Gemini: The AI agent is powered by Gemini, a large language model.
• Common Sense and Contextual Understanding: Gemini provides the agent with human-like common sense and the ability to understand context. This allows the agent to infer user needs based on situations (e.g., understanding what items might be suitable for a specific home office setup).
• Personalized Suggestions: The agent can suggest additional items that complement existing ones or fit a particular scenario, going beyond simple historical data.

The core difference lies in the intelligence of the AI agent and its ability to understand nuanced context, leading to more relevant and "generative" recommendations.

Getting Started with ADK

For developers interested in building similar AI agents, the demo points to the ADK GitHub page.

• Accessibility: The ADK can be found by searching "ADK GitHub."
• Tutorials: The page offers a range of tutorials to help users get started.
• Ease of Use: It's highlighted that users can create their first AI agent with just a few lines of Python code.

Conclusion

The "Surplus Cons" demo effectively illustrates the power of the ADK and Vati Vector Search in creating intelligent AI agents for retail. By leveraging Gemini's contextual understanding and advanced search capabilities like deep research and image analysis, these agents can provide a more personalized, efficient, and intuitive shopping experience compared to traditional recommendation systems. The availability of tutorials on GitHub makes it accessible for developers to explore and build their own AI-powered solutions.