New Framework for Integrating Generative AI into Traditional Systems

Integrating generative models into established computational systems offers immense opportunities but also carries significant risks, as many early adopters have discovered. This paper addresses these challenges by proposing a foundational framework for "Grounded Inference," designed to enable the deterministic encapsulation of probabilistic generative models within traditional system architectures. The framework introduces four specific primitives for AI-blended architecture, which are intended to facilitate a more predictable and controlled integration of AI components. Additionally, the manuscript identifies two prevalent anti-patterns observed across the industry, serving as crucial warnings for engineers working in this domain. The ultimate goal of this framework is to de-risk the incorporation of AI into traditional systems. It aims to provide a robust foundation upon which developers and generative model providers can build the next generation of reliable and predictable generative model interfaces.