Systematic analysis of operation and maintenance (O&M) data may play a significant role in reducing the cost of energy (CoE) of wind power. However, the maintenance work order text data remained comparatively underexplored due to the linguistic variability and complexity of unstructured text. While the applications of Large Language Models (LLMs) to automate manual maintenance-text analysis are emerging, smaller open-source models generally exhibited lower performance than larger proprietary models. In this study, a framework incorporating retrieval-augmented generation (RAG) is proposed to reduce the performance gap between small scale open-source LLMs and larger proprietary LLMs in wind turbine maintenance-text classification. Using a pre-labelled dataset, the proposed framework incorporates a compact, label-specific knowledge corpus containing class definitions, relevant terminology, and label boundaries to provide task-specific contextual information during inference. The findings show that the RAG-enhanced open-source model Qwen3.5-9B achieved a macro-F1 of 0.72, exceeding the 0.66 achieved by the baseline proprietary model GPT-5.6-luna, while remaining competitive with GPT-5.6-luna when both models were evaluated under the RAG configuration. Large inter-model variations observed in class-level F1-scores were also reduced for most label classes following the application of RAG. Despite the additional input-token requirements introduced by RAG, only marginal trade-off in inference throughputs was observed among the locally deployed models. These findings indicate that task-specific retrieval using a compact, label-structured knowledge corpus can improve the competitiveness of smaller open-source LLMs for wind-turbine maintenance work-order classification, although further evaluation across additional and larger maintenance datasets, retrieval corpora and LLMs is required to establish the generalisability and scalability of the proposed framework.