Researchers introduce LASAGNA, an open-source framework for controlled evaluation of Retrieval-Augmented Generation pipelines, testing 450 configurations across eight benchmarks with corpora ranging from 5,000 to 200,000 documents. They find that individually strong retrieval techniques — predicted to deliver a combined 47% gain — only produced a 17% gain when stacked together, and that optimizing for retrieval quality can actually hurt answer faithfulness. The best-performing configuration also changed as corpus size grew, leading the authors to conclude that isolated component gains are unreliable predictors of full-pipeline performance.
