Extracting Self-Consistent Causal Insights from Users Feedback with LLMs and In-context Learning