Perspectives on eCommerce Information Retrieval

Key Insight: Search as a Journey, Not a Single Event
Key Insight: The Difference Between Search and Agents
Key Insight: The Dynamic Nature of Search and Business Goals
Key Insight: The Reality of Embeddings and the Power of Multimodality
Key Insight: Balancing Exploration vs. Exploitation and Continuous Measurement
Q&A: The Role of Agentic Search in E-commerce
Q&A: Tracking Quality and Moving Beyond Basic Hybrid Search
Q&A: Defining a "North Star" Metric and the Importance of A/B Testing
Q&A: Layering Retrieval Techniques and Automating Relevance
Q&A: The "Query to Prod to Vec" Model Explained
Q&A: Top Advice for Implementing a New Search System
Q&A: Fine-Tuning Embeddings for Diverse B2B Domains
Final Summary: Search is a System, Not Just an Algorithm

Transcript

1. I’m ANDY <p>Co-Founder & CTO @searchhub.io</p> Deep passion for improving Product Discovery (Search & Recs) for over 2 decades.
2. a SEARCH-AGENT that optimizes your existing site search, at scale, without replacing it! WHAT IS ? enhancing >100 billion searches, generating around €5.5 billion in annual revenue for our clients. <p> visit searchhub.io </p>
3. HOW IT WORKS Full behavioral Attribution & Experimentation A Thin-Client for: ❏ Query Reformulation ❏ Query Understanding ❏ NeuralInfusion agentQuery
4. Are we all doing “retrieval” wrong ? <p> Practical lessons from >200 customer projects </p>
5. SEARCH IS A JOURNEY NOT A SINGLE EVENT OR FEATURE 01
6. SEARCH IS NOT AN AGENT SEARCH is great for exploration: It enables users to quickly & efficiently browse through & explore large sets of items. 02 While AGENTS are great for assistance: They help users to narrow down choice through guidance and incorporated feedback.
7. SEARCH IS HIGHLY DYNAMIC ❏ Indexed documents change over time ❏ Incoming queries change over time ❏ Configurations change over time (Syns, Boostings, Rewrites) ❏ Features change over time You face different VERSIONS of your SEARCH over time CONFIG INDEX QUERIES CONFIG CONFIG INDEX QUERIES CONFIG CONFIG INDEX QUERIES CONFIG CONFIG INDEX QUERIES CONFIG 03
8. 04 <p> The majority of ingredients for a successful search system constantly change over time. What we don't know, destroys our search softly. </p> Relevance Engagement Revenue Attractiveness Revenue EBIT vs. vs. vs. <p> a 10% increase in cost/query needs 100% growth in revenue for break/even @10% margin </p>
9. EMBEDDINGS ARE MESSY… 05 <p> In vector space search essentially transforms into navigation. </p> <p> Without clear borders or markers it’s almost impossible to efficiently navigate through it as distance != similarity </p>
10. 06 …UNTIL ADOPTED TO YOUR DOMAIN <p> By training or fine-tuning we manipulate the embeddings to form dense clusters with minimal semantic overlap. </p> <p> This is the only way to distinguish real semantic relationships, produce accurate and meaningful results. </p>
11. 07 YOU DON’T NEED LOTS OF DIMENSIONS <p> Our multi-modal Intent Embeddings with 128 dimensions. </p> <p> No need for GPUs & almost no measurable added latency. </p> <p> BUT maybe a bit less generalization capability. </p> inspired by Query2Product2Vec
12. FEEDBACK - <p> Tracking captures the what not the why and how. </p> <p> and most importantly not what could have been. </p> 08 EXPLORATION vs. EXPLOITATION
13. 20% of the assortment drives >87% of the overall product exposure Only exposed SKUs can generate Feedback like clicks, carts and buys SKU performance is heavily exploited Product-Exposure Product-Margin 09
14. DO CHANGES IMPROVE OR DECREASE SEARCH QUALITY OVER TIME? Local Optimum Global Optimum 10 <p> more than 85% don’t know for sure </p>
15. THANK YOU! Questions? find me here <p> visit searchhub.io </p>