Researchers have developed "GoldenRetriever," a novel framework for privacy-preserving retrieval-augmented generation (RAG) that utilizes non-interactive homomorphic encryption. This approach addresses the privacy concerns associated with RAG systems that typically process sensitive data in plaintext. By employing threshold selection instead of expensive top-k ranking under encryption, GoldenRetriever significantly reduces latency and computational complexity. The system uses CKKS-based homomorphic computation for encrypted similarity evaluation and document selection, ensuring that query content and intermediate scores remain confidential. Experiments show competitive retrieval effectiveness and a substantial reduction in latency compared to previous encrypted methods, paving the way for more secure and scalable RAG applications. AI
IMPACT Enhances privacy in RAG systems, potentially enabling wider adoption of LLMs with sensitive data.
RANK_REASON The cluster contains two arXiv papers detailing novel research on privacy-preserving retrieval methods for RAG systems.
- arXiv
- CKKS
- Dongfang Zhao
- Hugging Face
- Paillier
- GoldenRetriever
- homomorphic encryption
- large-language models
- private information retrieval
- retrieval-augmented generation
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
