Prompt Structure and Storage Architecture Mapping

This white paper outlines the relationship between prompt structure and storage architecture in the context of long-context inference and retrieval-augmented generation (RAG). It presents a mapping artifact that connects various layers of prompts to state objects, input/output patterns, and service level objectives (SLOs). The document is intended for infrastructure and platform teams tasked with designing and qualifying AI serving stacks. It emphasizes the importance of understanding how storage integrates into the serving path alongside compute, memory, and network resources. The paper discusses implications of long context and agentic workflows, which significantly increase state and I/O requirements. It also defines procurement success in terms of latency metrics rather than peak input/output operations per second (IOPS). The document includes a detailed examination of how prompt structure influences the performance and qualification of storage systems, providing a framework for optimizing AI infrastructure.