🧪 Test?View on arXiv
BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms
ragretrievalbenchmarkingenterprise-search
2607.26497
Builder Relevance
Jul 3195%
Abstract
A controlled study compares RAG paradigms over a 28-tier enterprise corpus ladder up to 511,959 documents and approximately 601M tokens.
Reality Card
Core Claim
At enterprise scale, lexical BM25 can outperform the evaluated file-system agent and dense-retrieval pipelines; the reported crossover occurs near 10M corpus tokens.
Method / Result
At the full evaluated corpus, the paper reports BM25 at 50.5 combined score versus 30.7 for its file-system agent configuration and 29.9 for DenseRAG.
Limitations
Results are tied to EnterpriseRAG-Bench, the selected reader, chunking, budgets, and evaluated implementations; they should not be treated as a universal retrieval ranking.