REFRAG proposes faster RAG with reusable embeddings

Published on September 15, 2025, the report presents REFRAG, a method that changes how retrieval-augmented generation (RAG) systems use retrieved content. Instead of supplying most retrieved tokens, the method replaces them with precomputed, reusable chunk embeddings.

According to the post, REFRAG achieves 30× higher speed and supports 16× longer contexts without loss of accuracy. These are claims made in the report; consult the original paper to verify evaluation conditions, technical details, and the comparison used. If you use AI to study or apply the work, avoid entering personal data or confidential documents without authorization.