ResearchResearch paperRetrieval, RAG & Search · Evaluation & Benchmarks2 sources · Oct 21, 2026

Retrieval Augmented Generation Compositionality

Retrieval Augmented Generation (RAG) grounds large language models in external knowledge.

Modern pipelines combine embedding, chunking, retrieval, reranking, query augmentation, context compression, and agentic verification, yet these techniques are typically evaluated one at a time.
It remains unclear whether their isolated gains survive composition, whether retrieval quality and answer faithfulness favor the same design choices, and whether conclusions from small corpora hold at large scale.
We pre-declare three hypotheses and introduce LASAGNA, an open-source framework for controlled RAG evaluation.
First, individually strong techniques from different phases of a RAG pipeline do not necessarily compose: three retrieval techniques predict a +47% combined improvement but deliver only +17% when used together.

[1]Retrieval Augmented Generation Compositionality
Hugging Face Daily Papers · Oct 21, 12:00 AM
Retrieval Augmented Generation (RAG) grounds large language models in external knowledge.
Modern pipelines combine embedding, chunking, retrieval, reranking, query augmentation, context compression, and agentic verification, yet these techniques are typically evaluated one at a time.
[2]Retrieval Augmented Generation Compositionality
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 01:30 PM · same content

Extractive summary: sentences quoted from the sources.

Before this