Reference Architecture for GenAI Applications — Part One
This past year has brought significant developments in the life of a machine learning engineer. Although I once regarded my article on the "Lifecycle Patterns of Generative AI" as somewhat dramatic, the progress of GenAI has surpassed all expectations.
We crafted prompt libraries, revamped workflows, integrated GenAI-backed features into existing products, and developed GenAI applications. As we navigated through various challenges, my teams consistently encountered similar observations, recurring issues, and repeatable solutions. These patterns pointed unmistakably toward the necessity for an architecture blueprint tailored for GenAI applications. While we came across numerous insightful papers, articles, and blog posts, only a handful dared to tackle the entire stack or landscape.
Utilizing these resources proved invaluable by offering a shared vocabulary, framework, and an extensive solution stack to anchor our discussions. They empowered our teams to develop a clear solution architecture hypothesis and identify the tools, products, or services to integrate into either the solution design or delivery.
We've customized the GenAI applications reference architecture to align with our experiments and implementations. Here are some insights on what we adjusted and the reasons behind our modifications:
1. Below arrangement of blocks has been perfected to better capture our perspective on the GenAI applications landscape.
2. The diagram may appear overwhelming, and we apologize for that. However, consolidating everything in one place has significantly aided our understanding and facilitated the development of various compelling GenAI solutions. We trust you'll recognize the benefits of this approach.
3. Some components on the diagram may lack industry traction, but we've included them with the intention of piloting ideas or posing questions. For instance, Prompt Validation (pre-submission) is a concept we haven't found tools or GitHub projects for, yet it seems a meaningful addition to the responsible AI and RBAC tool kits.
4. The products/tools/platforms within dotted rectangles aren't meant to be exhaustive; they represent those we've either used or experimented with. We'll continue to update and refine these rectangles as we become acquainted with more available options.
5. The GenAI landscape is rapidly evolving, and our diagram may have become outdated shortly after its creation. Our understanding is subject to limitations and may differ from yours.
Reference Architecture for GenAI Applications
The second part of this post will explore several solution archetypes and illustrate how they can be represented using the reference architecture.