$\pi$-RAG: Oblivious Retrieval via Semantic Quantization and Transcendental Addressing for Large Language Models

The paper presents $\pi$-RAG, a new architecture designed for oblivious retrieval in Large Language Models (LLMs) that mitigates risks associated with sensitive data exposure. It employs the digits of $\pi$ to create a transcendental addressing mechanism, which serves as an immutable layer between the LLM and private data, ensuring semantic understanding while maintaining privacy. The architecture incorporates a Semantic Quantization Layer that maps user inputs to Canonical Intent Centroids, enhancing security and compliance for applications in high-stakes sectors like finance and healthcare.