Computing with tensors in Vespa

Vespa introduces a unique tensor formalism that differs from conventional machine learning frameworks by using named dimensions and supporting both sparse and dense tensor types, alongside a concise set of core functions. This approach enhances tensor computation by allowing semantic verification and static type inference, thus reducing runtime errors. Vespa's tensor framework employs a small set of foundational operations—such as map, reduce, and join—enabling complex computations without extensive libraries, which improves interoperability and maintainability. The framework supports dense, sparse, and mixed tensors, allowing flexible data representation and efficient space utilization. By leveraging named dimensions, Vespa provides a robust type system that simplifies tensor operations, making numerical computation more expressive and easier to optimize. The recently introduced tensor playground offers an interactive environment for exploring these tensors and tensor expressions.