From Source Code to Kubernetes Manifests: Guiding LLMs with Structured Metadata

Deploying applications on Kubernetes (K8s) often requires writing and maintaining complex manifests that accurately reflect the application’s architecture and requirements. This task requires significant expertise, time, and effort, especially for heterogeneous architectures composed of multiple microservices, interdependencies, and configuration options. In recent years, Large Language Models (LLMs) have been proposed to automate cloud application lifecycle phases, including manifest generation. However, obtaining accurate manifests requires specific domain knowledge about both the application and K8s concepts, which can be difficult to convey to LLMs effectively. Attempting to address this by feeding the entire codebase quickly exceeds the model’s context window, while using simple natural language instructions often leads to incorrect or incomplete manifests. Filling this gap, we present a novel pipeline that leverages structured metadata to guide LLMs in generating accurate and deployable K8s manifests. Our approach directly scans application repositories, deterministically extracting metadata about each microservice and its dependencies. This information is structured into an Intermediate Representation (IR), a concise summary of the application’s architecture and requirements, which is then used to prompt LLMs for manifest generation. We evaluate our pipeline on a dataset of 40 microservice architectures, demonstrating significant improvements in deployment success rates: from 15% with direct prompting to 47.5% with IR guidance, and up to 62.5% when incorporating minor manual overrides. While the results highlight the effectiveness of structured metadata in enhancing LLMs’ performance, they also underscore the need for human guidance when dealing with complex and heterogeneous scenarios.