Privacy-Preserving RAG via Multi-Agent Semantic Rewriting: Achieving Confidentiality Without Compromising Contextual Fidelity

The article presents a multi-agent framework for privacy-preserving Retrieval-Augmented Generation (RAG) that sanitizes retrieved content through semantic rewriting, effectively mitigating privacy leakage in sensitive applications. The framework employs three specialized agents for privacy extraction, semantic analysis, and reconstruction, achieving a reduction in targeted information exposure from 144 instances to just 1 in the LLaMA-3-8B model, while maintaining a BLEU-1 score of 0.122, surpassing the SAGE method. This approach introduces no additional latency for online inference, as the rewriting occurs in a one-time offline preprocessing step, making it practical for deployment in real-world scenarios.