A Hybrid Approach for Intent-Driven Recommendation of CEI Orchestration Frameworks

Abstract

Modern IT applications span heterogeneous environments, from single-cloud and multi-cloud infrastructures to cloud–edge and cloud–edge–IoT (CEI) continuums, and involve diverse resources and services requiring distinct orchestration capabilities. This makes the selection of orchestration frameworks that automate the coordination and management of services and resources, particularly challenging. Existing approaches mostly rely on manual analysis and offer limited support for automatically mapping application requirements to orchestration capabilities. While large language models (LLMs) can interpret natural-language user intents, their standalone use for orchestration framework recommendation raises concerns about reliability and the absence of explicit reasoning. To address these challenges, we propose an intent-driven approach for orchestration framework recommendation. It combines LLM-based natural-language interpretation with grounding in a verifiable and reliable Knowledge Graph (KG) that captures orchestration frameworks, their capabilities, and their distinguishing features. In addition, rather than relying on a purely generative process, it decomposes recommendation into three distinct steps: interpreting natural-language requirements into a structured intent representation, identifying candidate orchestration frameworks from the KG, and analyzing their suitability to determine whether the user requirements are best satisfied by a single framework, a combination of frameworks, or only partially covered, while providing transparent and intent-aligned justifications. Experimental results, complemented by a user study, confirm that LLM-Grounded KG with an explicit decomposition of the recommendation process significantly improves accuracy, transparency, and overall recommendation quality compared with LLM-only approaches. Moreover, our results show that integrating smaller language models (SLMs) into our approach yields better performance than using standalone LLMs.