IBM's LASAGNA study shows RAG technique gains shrink when combined: +47% predicted becomes +17% realized

ibm · hf · 2026-10-22

IBM released the LASAGNA study on Hugging Face, evaluating composed RAG pipelines across corpus scales.

Key finding: gains measured for techniques in isolation do not reliably predict combined retrieval quality (NDCG@10), faithfulness, or large-corpus performance
Gains shrink when combined: NDCG@10 predicted +47% vs. +17% realized; faithfulness predicted +28% vs. +32% realized — and the loss depends on the metric
Decay curves diverge sharply as corpora grow from 5K to 200K+ documents (e.g., a Speed config shows steep decline with R²=0.90, Max Reliable only R²=0.30)
Practical implication: you can't extrapolate single-technique benchmarks to full RAG pipelines on large corpora; composition-level evaluation is required

More from Research