Abstract:
The rapid growth of the Flutter ecosystem has built significant challenges for the developers
in choosing the right third-party packages. Existing popularity-based methods fail to capture
developer intent or semantic alignment with real-world requirements. To address this gap, this
research presents a hybrid LLM-based recommendation system that integrates structured metadata
from Flutter repositories with unstructured knowledge from Stack Overflow. With three
LLMs comparatively evaluated, GPT-4- Turbo was chosen based on the accuracy, consistency,
and low hallucination rate. The final recommendation model uses a weighted scoring method
that integrates popularity, recency, likes, downloads, accepted answers and LLM-based semantic
relevance, which generates ranked package recommendations based on the user’s requirements.
Experimental evaluation across multiple test scenarios achieved a precision of 93.2%,
recall of 91.5%, F1-score of 92.3%, and an average semantic relevance score of 94/100. Expert
evaluation by Flutter practitioners with six to ten years of experience confirmed that 92% of the
recommendations aligned with realworld development requirements. Even though the paper
focuses on state management packages, the approach methodology can be generalized to other
Flutter package categories. The results highlight the potential of LLMbased hybrid systems to
improve developer decision-making.
