Enhancing Flutter App Development: Addressing Configuration and Compatibility Bugs Using Large Language Models

Abstract

In today’s world, there are cross-platform frameworks available, like Flutter, that allow the developer to develop an app from a single codebase. However, debugging issues and compatibility with different devices and dependencies are still the greatest challenges for Flutter developers because there are continuous framework changes and divergent platform-specific behaviours. To overcome these challenges, the research adopts the use of advanced LLMs, including GPT-4o, Claude Sonnet 3.5, and Gemini 2.0 Flash, combined with RAG capabilities. The collection of data from multiple sources like Stack Overflow, GitHub, and Flutter documentation initiated the study. The dataset was cleaned and preprocessed using markdown/HTML stripping, deduplication, code-text separation, and case normalization. To kick off the evaluation a fixed set of prompt templates to 350 Stack Overflow questions across all the LLMs were applied. The assessment consisted of two approaches: expert validation of 25 key questions and assessing high-scored Stack Overflow answers using cosine similarity and prediction correlation against the developed framework and standalone LLM. Through this assessment, it was able to gauge the models’ efficacy in finding configuration bugs and addressing them. The RAG pipeline, designed with a dual-model embedding approach (separate models for text and code) in a vector database, achieved a higher similarity score of 0.8371. Retrieval performance was further improved through re-ranking methods. Early evidence hints at the enhanced accuracy/precision of a RAG-enhanced LLM when that LLM is a standalone model for better Flutter configuration. This paper presents a modular framework to integrate RAG-enhanced LLMs for bug detection and fixing. The framework requires much lesser debugging efforts along with more reliability of the app. Additionally, it can also adopt the fast-evolving Flutter framework. By filling a significant gap in cross-platform development literature, it helps advance AI-assisted debugging and improve the development workflows of Flutter apps.