[DSC DACH 24] Enhancing E-commerce Search with Custom Embeddings and Generative AI - Stanko Kuveljic

This talk delves into the innovative use of custom embeddings and re-rankings to enhance the e-commerce search experience. It will explore the necessity of rerankers and the limitations of embedding models, and how incorporating user history and engagement can significantly improve search results. Attendees will gain insights into fine-tuning and training custom embeddings using historical engagement data. Additionally, the talk will cover the application of generative AI to synthetically produce datasets, and discuss future steps, including the integration of product metadata and images. This session is perfect for anyone interested in the cutting-edge intersection of AI and e-commerce.

1. Enhancing E-commerce Search
with Custom Embeddings and
Generative AI
Stanko Kuveljić

2. E-Commerce Search

3. Traditional Keyword Search

4. Semantic Search

5. How Does Semantic
Search Work?

6. Embeddings

7. Vector Search

8. …. BUT

9. Limitations of Vector Search
• Limited interpretability
• Query bias and sensitivity
• Misses specific details and data
• Struggles with complex queries
• Lacks user engagement context
• Long tail challenges
• Search relevancy issues

10. Agenda
• Semantic Search vs Reranking
• Bi-Encoders
• Cross-Encoders
• Cold Start - Data Generation
• Context Enrichment

11. Semantic Search
vs Rerank

12. Semantic Search
vs Ranking

13. Bi-Encoders

14. Bi-Encoder Architecture

15. Training Bi-Encoder - Triplet Loss

16. Cross-Encoders

17. Cross-Encoder Architecture

18. Training Cross-Encoder

19. Which one is better?

20. Bi-Encoders vs Cross Encoders

21. What if there are no
queries?

22. Data Generation

23. Untapped Potential

24. Context Enrichment

25. Thank You!