Reproduce “RAG Components Are Not Independently Rankable” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vishnu Prasad Vijaya Kumar, Iaroslav Neverov, Pavel Chizhov, Anton Changalidis, Vaiva Kazanaviciute, Pierre-Carl Langlais, Anastasia Stasenko, Ivan P. Yamshchikov
The authors evaluate retrieval-augmented generation evidence delivery by swapping components while holding questions and visible token budgets fixed. They measure claim recall: the fraction of reference claims supported by retrieved context. On FinanceBench, Semantic-512 changes recall by +0.27 points with sparse retrieval, −2.59 with dense retrieval, and +2.65 with late interaction. These results cover five tested benchmarks, not answer accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.