This study explores how GPT-3.5 Turbo can be enhanced to improve its predictive performance and explainability from social media posts. Two strategies are employed, first by implementing retrieval augmented generation (RAG), where relevant clinical knowledge from the NICE guidelines is integrated into GPT-3.5 Turbo to enhance its reasoning in depression detection. The second involved using prompt techniques, zero-shot and few-shot, guiding the model’s understanding of the task through examples of depressive and non-depressive content from the social media data. These were evaluated on standard performance metrics, including accuracy, precision, F1-score, and recall. Assessment of the explanations of the model configurations took a non-clinical qualitative approach, measured by the clarity of reasoning, references made to specific guidelines and the consistency in explanation. The findings revealed that incorporating the clinical guideline improved the performance of the baseline GPT-3.5 turbo in the zero-shot configurations. However, when combined with few-shot, the RAG-enhanced models tend to produce fewer positive predictions for depressive posts, raising concerns in clinical applications. In contrast, few-shot configurations without RAG demonstrated stronger balance across metrics.