Leveraging Large Language Models for Intent-Based Generation of Cloud-Native Configurations

Cloud-native systems have revolutionized the development, deployment, and management of modern applications, enabling organizations to achieve unprecedented scalability and resilience. However, the complexity of configuring and managing these systems remains a significant barrier, particularly for users lacking deep technical expertise. In this paper, we present an intent-based framework that leverages large language models (LLMs) to automate the generation of configurations within cloud-native environments, focusing on Kubernetes deployment workloads as a use case. The framework offers a practical toolkit for users who lack deep technical expertise, helping them manage Kubernetes configurations by converting user intents into workable, deployable configurations. Our evaluations demonstrate that our proposed framework achieves a 100% success rate in schema validation through kubectl using few-shot prompting, compared to $\mathbf{9 7. 1 8 \%}$ with zero-shot prompting and $\mathbf{9 5. 7 7} \%$ with human-generated configurations. In semantic validation using Kubeval, our framework performed well, achieving a success rate of $\mathbf{9 1. 5 5 \%}$ with both few-shot and zero-shot prompting, compared to 87.32% with human-generated configurations. These results highlight our framework’s potential to enhance both the accessibility and efficiency of cloud-native technologies, making complex deployment tasks more approachable for a broader range of users.