In low-resource framework software development (e.g., HarmonyOS), large language models (LLMs) typically lack exposure during pre-training, leading to poor performance when generating code for such frameworks. While LLMs often preserve code logic across different programming languages, they tend to fail on framework- specific APIs and syntax errors. This suggests that pre-training enables LLMs to master general algorithms, but they remain unfamiliar with the syntax and API characteristics of unseen low-resource languages or frameworks. Consequently, even large-scale commercial models such as GPT-4o struggle to produce correct code without prior knowledge of these elements. Inspired by these challenges, we introduce APIKG4Syn, a framework that leverages API knowledge graphs to generate API-Oriented question–code pairs without requiring executable code for low-resource framework. APIKG4Syn provides both single-API information and multi-API information, the latter identified through uncertainty estimation (UE)-guided Monte Carlo Tree Search (MCTS), to construct a comprehensive dataset for fine-tuning LLMs in low-resource scenarios. To evaluate APIKG4Syn, we select HarmonyOS as a case study and develop the first HarmonyOS benchmark for code generation. Experimental results demonstrate that Qwen fine-tuned with APIKG4Syn achieves a pass@1 of 25.00%, surpassing the untuned GPT baseline at 17.59%. These findings underscore the effectiveness of API-Oriented data in improving LLM performance for low-resource software development.