How the paper used Lasagna

Lasagna is the framework behind the paper Retrieval Augmented Generation Compositionality by Ido Levy, Segev Shlomov and Leshem Choshen (2026). This page says what the study asked and how the framework made it possible. For the full results, read the paper.

The question

RAG papers report one component at a time. A new reranker is measured against a baseline with no query expansion and an old chunker. If a reranker claims +4 NDCG and a query expander claims +3, you cannot tell from the papers whether stacking them gives +7, +4 or -1.

The study asks whether gains measured one component at a time survive when the components are combined.

The design

The study tests 450 configurations across eight benchmarks, with corpora from 5K to 200K documents. Eight design dimensions vary across the configurations.

To attribute a difference to a single change, every other part of the pipeline has to stay fixed. That needs a pipeline with swappable stages and a harness that checks the swap. Lasagna provides both: the stages, the comparability key and the significance tests.

What it found

Finding	Predicted	Realized
Retrieval gains do not add up: the three strongest single techniques, combined (NDCG@10)	+47%	+17%
Faithfulness gains do carry over: one factor at a time versus the best full pipeline	+28%	+32%

The best pipeline also changes with corpus size. At 200K documents, the design that leads at small corpora trails by 13.7% NDCG@10, and the rank flip is visible from 25K.

What the paper calls the profiles

The paper's Max-Reliable profile is max_quality in the code. See profiles and configuration.

Run your own version

You can ask the same question about your own components. Run the arms alone and together, then use lasagna results interaction on the 2x2. The compare a component guide shows how to write the sweep.

Lasagna does not ship the paper's per-query results or a prebuilt index. The pre-registration has the study's plan.