The article proposes a new framework, LongRAG, that aims to improve the performance of Retrieval-Augmented Generation (RAG) by using long retriever and reader components. LongRAG processes Wikipedia into larger 4K-token units, reducing the total units from 22M to 600K, thus decreasing the burden on the retriever. The top-k retrieved units (≈30K tokens) are then fed to a long-context Language Model for zero-shot answer extraction. LongRAG achieves EM of 62.7% on NQ and 64.3% on HotpotQA (full-wiki), which is on par with the state-of-the-art model.
