wardcat: On-Premise Data Privacy Architecture for RAG and LLM Projects

On-premise data privacy architecture named wardcat was created by Oguz Topcu to address security risks and data leaks in Retrieval-Augmented Generation (RAG) and large language model (LLM) projects. This architecture, published on PyPI, goes beyond basic pattern matching by combining Regex, SpaCy Named Entity Recognition (NER), and local open-source large language models like Qwen3:14b.

The hybrid layered approach enables accurate detection of sensitive data, with texts being anonymized through reversible masking and restored to their original form post-LLM processing. The wardcat library also features an is_sensitive() function that applies a contextual security filter to prompts before they are processed, enhancing protection against data leaks.

A comparison with Microsoft Presidio, currently considered the industry standard, reveals that wardcat's rule-based (Regex) and linguistic (NER) layers outperform Presidio in terms of the F1 score, which measures the accuracy of the model in detecting sensitive information. When the LLM layer is added, wardcat achieves an impressive F1 score of 0.920 in challenging scenarios, surpassing Presidio's 0.838, and 0.875 in independent corpora like Gretel Finance.

Moreover, the is_sensitive() filter outperforms traditional rule engines, achieving 88% accuracy in sensitivity classification, compared to Presidio's 63%. These results highlight the importance of adopting a hybrid, layered architecture to safeguard on-premise pipelines from data leaks. For those looking to implement wardcat, the project can be found on GitHub, its documentation can be accessed at docs.wardcat.com, and it is available on PyPI at https://pypi.org/project/wardcat/.