This paper introduces $Urania$, a novel framework for generating insights into large-scale language model (LLM) chatbot interactions with strict differential privacy (DP) guarantees. $Urania$ employs a privacy-preserving clustering mechanism and innovative keyword extraction methods, including frequency-based, TF-IDF-based, and LLM-based approaches. Leveraging DP tools such as clustering, partition selection, and histogram-based summarization, $Urania$ provides end-to-end privacy. We evaluate lexical and semantic content preservation, pairwise similarity, and LLM-based metrics compared to a non-privacy-preserving Clio-based pipeline (Tamkin et al., 2024). We also develop a simple empirical privacy evaluation demonstrating the enhanced robustness of the DP pipeline. The results demonstrate that the framework effectively balances data utility and privacy by extracting meaningful conversational insights while maintaining strict user privacy.
