Retrieval-augmented generation (RAG) lets large language models pull in external structured knowledge when answering queries. The approach can improve accuracy and relevance, but it also raises privacy concerns because the retrieved information may include sensitive content.
A new arXiv preprint, "Private Retrieval Augmented Generation via Random Projection," addresses this issue. The authors argue that RAG can leak sensitive information from the knowledge source, and they propose a method based on random projection to mitigate the risk. The abstract does not give experimental details, so the method's effectiveness is not yet demonstrated.
The paper is available on arXiv under identifier 2610.22297. As with many preprint proposals, further evaluation and peer review will be needed to determine whether the approach works in practice.