Design and Evaluation of a Privacy-Preserving Retrieval-Augmented Generation Framework for Scalable Group Messaging Systems

This paper presents a local implementation of the Retrieval-Augmented Generation (RAG) model for private group messaging systems that require scalability, long-term conversation history storage, and low retrieval latency.The paper explains the integration of the RAG model with query decomposition, multi-hop retrieval, and hybrid reranking to support context-aware reasoning over chat history without external services.Performance evaluation showed retrieval latency below 100 ms for 100 concurrent queries when processing 50,000 stored messages on multiple hardware platforms.The study compared answers from RAG-based models and a non-RAG model using 150 human-verified QA pairs.All RAG-based models produced factual answers with a faithfulness score of at least 0.8, and the Qwen-3 (8B) model achieved the highest semantic accuracy of 0.86.The results indicate a significant difference in scores between the non-RAG and RAG models, demonstrating that retrieval and indexing capabilities are necessary for highly accurate and faithful responses.