AssistantGraph: An Approach for Reusable and Composable Data-Driven Assistant Components

Steady progress in ubiquitous technologies and machine learning facilitates ever-new and better digital assistants. However, most of these emerging assistants rely on - partly similar - data-driven analyses that are independent of each other, leading to redundancy issues. In this paper, we propose a novel concept (termed AssistantGraph) for an efficient design and runtime support of digital assistants. More specifically, assistants need to represent their data-driven processing pipelines as a directed acyclic graph of assistant components (modularity) to benefit from serverless computing with data access. Facilitated component sharing across assistants (reusability) leads to a more connected and efficient overall graph: a shared component instance requires to run only once; versioned components are enabled by the proposed backward chains of converters (versioning). We further develop data and control flow mechanisms through recursive filtering on demand to trigger the data-driven components as required. Within a novel proposed and prototypically implemented open assistant infrastructure, we evaluate our concept in terms of feasibility and performance. The results show reduced redundancy with simultaneous significant performance gains (through component sharing) despite minimal additional overhead (due to modularization and backward compatibility). We believe that our approach gives a new perspective on data-driven assistants and complements an open assistant ecosystem.