"RemoteRAG: A Privacy-Preserving LLM Cloud RAG Service"

RemoteRAG enables private cloud-based RAG services by protecting user queries while maintaining 100% retrieval accuracy through a novel privacy mechanism and efficient search range limitation.

-----

https://arxiv.org/abs/2412.12775

🤔 Original Problem:

Cloud-based RAG services require users to send queries in plaintext, exposing sensitive information like health conditions or financial status.

-----

🔧 Solution in this Paper:

→ RemoteRAG introduces (n,ε)-DistanceDP, a privacy mechanism that controls query privacy leakage using a budget ε in n-dimensional embedding space.

→ It limits search scope from total documents to a smaller set related to perturbed embeddings, drastically reducing computation costs.

→ The system offers dual retrieval methods - direct retrieval or k-out-of-k' oblivious transfer protocol based on privacy requirements.

→ A theoretical framework ensures minimum search range size contains target documents while optimizing efficiency.

-----

💡 Key Insights:

→ Privacy and efficiency can coexist in RAG systems without sacrificing accuracy

→ Perturbed embeddings effectively protect query privacy while maintaining retrieval quality

→ Search range limitation significantly reduces computational overhead

-----

📊 Results:

→ Achieves 100% retrieval accuracy across various document counts and embedding models

→ Processing time: 0.67 seconds vs 2.72 hours in non-optimized privacy schemes

→ Data transfer: 46.66 KB vs 1.43 GB in baseline privacy-conscious approaches