Skip to content

Summary

This completes the Vector Search hands-on. Great work! 🍺

What You Learned

Value of Building Blocks + IBM Bob

  • Significant development time reduction: Completed in approximately 90 minutes what would take days to weeks
  • High-quality implementation: Code generation based on best practices
  • Natural language instructions: Feature addition possible without programming knowledge

Implemented Features

  1. Product image display: Added images to search results
  2. Price filter: Filter by price range
  3. Recommendation reason: Display why a product is recommended

Vector Search Overview

  • Searches by understanding the "meaning" of words
  • Unlike traditional character search, finds similar meanings even with different phrasing
Use Cases in IBM Products
  • watsonx Discovery: Automatically extracts relevant information from large volumes of documents and presents optimal answers to user questions (product image display)
  • watsonx Assistant: Understands customer inquiries and automatically generates optimal responses from similar past cases (price filter)
  • watsonx Orchestrate: Understands entire business processes and automatically executes appropriate workflows according to user intent (recommendation reason)
Use Cases in Familiar Services
  • Google Search: Finds pages and answer candidates with similar meaning even when the words do not exactly match (recommendation reason)
  • Amazon: Finds products by description or use case even when the product name is unknown, and recommends similar products (product image display, price filter, recommendation reason)
  • YouTube: Suggests videos users may want to watch next based on viewing history and video similarity (product image display, recommendation reason)
  • Netflix: Recommends titles close to watched genres, atmosphere, and viewing patterns (product image display, recommendation reason)
  • Spotify: Finds music close to favorite songs or playlists and reflects it in recommendations and automatic playlists (product image display, recommendation reason)
  • Instagram: Ranks posts and ads based on similarity in photos, videos, hashtags, and interests (product image display, recommendation reason)
  • Facebook: Shows feeds, groups, and ads close to post content and user interests (product image display, recommendation reason)
  • TikTok: Recommends short videos close to user preferences based on watching, skipping, and likes (product image display, recommendation reason)
  • Google Photos / Apple Photos: Finds photos with meaning close to words such as "sea", "dog", or "sunset" (product image display)
  • ChatGPT / AI Chat: Searches internal documents or knowledge close to a question and uses them as answer evidence (recommendation reason)

Deployment to Production Environment

Current Configuration (For Learning)

  • Hugging Face + Milvus: Completely free, offline support, optimal for learning

Migration to IBM Products

  • watsonx.ai: Enterprise-grade, advanced models, commercial support
  • watsonx.data: Large-scale data integration, governance features, petabyte support

Selection Guide

Scale Recommended Configuration
Learning/PoC Hugging Face + Milvus
Small-scale production Hugging Face + Milvus
Medium-scale production watsonx.ai + Milvus
Large-scale production watsonx.ai + watsonx.data

Value in Customer Systems

When integrating Vector Search into a customer's existing system, it is not enough to build only a search API. You need to connect data integration, embedding generation, vector databases, search APIs, screen display, and operations design. By using Vector Search Builder + IBM Bob, teams can reuse the foundation for technology selection and implementation while focusing on customer-specific requirements.

When integrating into a customer system without Vector Search Builder:

Integrating into a customer system without Vector Search Builder

When integrating into a customer system with Vector Search Builder + IBM Bob:

Integrating into a customer system with Vector Search Builder + IBM Bob

This difference makes it easier to deliver the following value in projects.

  • Faster startup: Prepare the basic Vector Search configuration in a short time
  • Focus on customer requirements: Spend time on differentiating parts such as business data, screens, search conditions, and explanation text
  • Easier iteration: Quickly tune search results and UI by asking IBM Bob in natural language

Challenge

Advanced Challenge: Comparison of Search Methods in Agentic RAG

Theme: Investigate the differences in Harness Engineering between Lexical Search and Vector Search in Agentic RAG!

Investigation Points

  1. Lexical Search

    • Traditional search like keyword matching, BM25
    • Search accuracy through exact or partial matching
    • Use cases in Agentic RAG

  2. Vector Search

    • Search based on semantic similarity
    • Representation learning through embedding models
    • Use cases in Agentic RAG

  3. Harness Engineering

    • How to combine both methods
    • Hybrid search implementation patterns
    • Scoring and re-ranking strategies

  4. Differences in Agentic RAG

    • Impact on agent decision-making
    • Relationship between search accuracy and response quality
    • Cost and performance trade-offs

Recommended Approach

  • Implement and compare both methods in actual use cases
  • Quantitatively evaluate search accuracy, response time, and cost
  • Utilize Building Blocks' Agent Builder mode
  • Document and share evaluation results

Through this challenge, you can understand important decision points in RAG system design.

Reference Materials