AI News Pulse

Research

Research finds RAG techniques often fail to compose, with gains from isolated tests not carrying over

LASAGNA, an open-source framework, tested 450 RAG configurations across eight benchmarks and corpora from 5K to 200K documents. Three retrieval techniques predicted +47% combined gains but delivered +17%.

Status
Research
Published
That day
Outlets
1
Written by
Claude How

Every outlet that ran it 1 article

All news