Natural Language Search on Semi-Structured Data

Описание: In this talk, we hear from Daniel Svonava, CEO at Superlinked, about why traditional search systems are fundamentally broken and how we can fix them. We explore why relying solely on text embeddings for search is problematic when dealing with numerical data, location information, and user preferences. He explains his "mixture of encoders" approach that combines specialized encoders to handle different data types more effectively than traditional methods.

Daniel also shares why re-ranking is often just a hack to fix poor retrieval, why filters are overused, and why the world can't be seen as just strings. We also discuss how this approach applies to real-world examples like hotel searches and how it compares to recommendation systems.

Daniel Svonava is the CEO & co-founder of Superlinked.com, an open source framework and ML infrastructure platform for building intelligent search, personalization and analytics experiences with vector embeddings that combine structured and unstructured data. Previously, Daniel was an ML Tech Lead at YouTube, where he built ads forecasting infrastructure that powered the buying flows of $10B/y-worth of ads.

If you want to learn more about improving rag applications check out: https://improvingrag.com/

TIME STAMPS

00:00 Introduction to Recommendation and Retrieval Systems
01:48 Understanding User Queries and Context
07:30 Issues with Current Search Approaches
14:33 Building Effective Search Systems
21:58 Case Studies and Real-World Applications
24:47 Open Source Framework and Cloud Product
27:24 The Future of Text Encoders
29:42 Superlink's Hotel Query Example
33:12 Fine-Tuning Text Embeddings
35:16 Refreshing Embedding Models
43:19 Graph Structures in Embedding Systems
45:23 Challenges with Text-Based Models
50:40 Conclusion and Final Thoughts