Context
An original research preprint evaluates 450 retrieval-augmented generation pipeline configurations across eight benchmarks and corpora containing 5,000 to 200,000 documents. Its authors report that component gains do not necessarily compose and that retrieval and answer faithfulness can favor different choices.
What changed
The LASAGNA framework supplies a controlled full-pipeline test of combinations of embedding, chunking, retrieval, reranking, augmentation and other RAG techniques; three retrieval methods whose individual gains suggested +47% improvement delivered +17% in combination.
Why it matters
AI discovery and source-selection systems need end-to-end retrieval and faithfulness measurements rather than relying on individual component benchmarks or extrapolating citation performance from retrieval accuracy alone.
Visibility OS interpretation
Visibility OS interpretation: Pipeline interactions can change which evidence reaches an answer, so isolated retrieval improvements are not reliable predictors of downstream answer grounding. This is not evidence about deployed commercial AI-search ranking.
What remains uncertain
Do the reported interaction effects replicate with other corpora, models and current production-scale web retrieval? Which combinations improve both retrieval and answer faithfulness? Do they affect actual citation selection?