Retrieval Augmented Generation Compositionality
Modern RAG research typically measures chunking, reranking, and query rewriting in isolation, assuming component improvements compound additively. Introducing the LASAGNA evaluation framework across 450 configurations and corpora up to 200,000 documents, researchers find that modular gains fail to stack predictably: three retrieval techniques projecting a +47% combined improvement yielded only +17% in practice. Crucially, optimizing for retrieval quality frequently trades off answer faithfulness, and configurations winning on compact benchmarks degrade as corpus scale grows, underscoring the necessity of testing pipelines as unified systems.
Why it's worth reading
Offers empirical proof across 450 pipeline configurations that isolated RAG improvements do not reliably compound and can actively harm generation faithfulness.