Building SOTA Search: It's Ranking All the Way Down

Effective Ranking consists of multi-dimensional functions that encode the full topology of a corpus and its query stream. Relevance is just one axis—others include trustworthiness, recency, diversity, personalization, and even negative constraints like safety or policy override. But how can we collect these various signals and how can we merge them all together at runtime? This talk will walk through what such an architecture looks like (inspired by decades of work on Google Search) and how to build it out in practice. It also covers how increasingly, many emerging AI challenges—prompt optimization, ensemble selection, synthetic data generation, reward modeling—are ultimately Search problems in disguise as they all consist of generating candidates, scoring them across multiple axes, and selecting the optimal ones. Recognizing this reframing allows us to bring the tools and rigor of information retrieval to problems far beyond traditional search.