Session Description

Query Understanding is one of the most important, yet usually overlooked, steps in information retrieval. Most search teams focus entirely on “algorithmic ranking”, hoping that the math (BM25, Vector Similarity, Hybrid Fusion) will magically rank the best results without first contextually understanding the queries. It often doesn’t.

In this talk, we walk through why Query Understanding is critical for a search pipeline, and why ranking alone is usually insufficient for optimal search. We’ll then show off a Query Understanding implementation using modern techniques like Wormhole Vectors (hopping back and forth between sparse and dense vector spaces) and late interaction (multivector representations per token) to significantly improve what was historically possible with more traditional query understanding approaches (like a knowledge graph + a text tagger).

The improvements aren’t just theoretical. We’ll show code examples (on several datasets) for interpreting query intent and rewriting queries prior to the final search request. We’ll then walk through our recent benchmarks of this approach that blow away today’s popular hybrid search techniques.