From BM25 to Mixture-of-Encoders - Evaluations for Next-Gen Search and Retrieval Systems

Explore the evolution of search and retrieval systems in this 48-minute conference talk that examines the limitations of traditional search methods when handling modern user queries requiring both structured and unstructured data. Compare the performance of keyword, vector, hybrid, and late-interaction models against Superlinked's innovative mixture-of-encoders approach through comprehensive retrieval evaluation methodologies. Learn how different search approaches handle real-world scenarios such as queries for "5 guests under $200 with 4.8+ rating" using benchmark datasets and production use cases. Discover Superlinked's mixture-of-encoders system that combines dedicated encoders for various data types including text, numbers, and categories with LLM-driven query understanding to enable more accurate and scalable retrieval. Gain insights into evaluation metrics, methodology best practices, and common pitfalls in retrieval system assessment. Understand how to productionize advanced search systems and explore practical applications across industries from travel to e-commerce, while looking toward the future of multi-attribute and metadata-aware embeddings search.