EnterpriseRAG-Bench: A RAG Benchmark for Company Internal Knowledge

A RAG benchmark for the real world.

The first RAG benchmark built on company-internal data, not Wikipedia. 500K+ docs, 500 questions, 9 enterprise sources. Open-source, MIT-licensed. By Onyx.

Put your RAG to the test → HuggingFace Leaderboard ↗

How does your RAG stack up?

Bars: per-metric scores on 500 enterprise RAG questions (higher is better).

Score % 0% 25% 50% 75% 100%
OpenClaw 68.2 61.0 50.2 49.0 48.4
OpenAI File Search 41.9 37.7 35.6 34.5 27.2
RAGFlow 25.0 24.9 24.9 24.9 24.9

ENTERPRISE RAG-BENCH · 500 QUESTIONS · 12 SYSTEMS
ONYX.APP/ENTERPRISE RAG-BENCH ↗

Inside the dataset

We generated a realistic synthetic company with documents across 9 different sources.

How it's built

  1. Scaffold
  2. Generate
  3. Add noise

Scaffold

Lay down the synthetic company. Each source is a window into the same world — same projects, same people, same decisions.
methodology.md ↗

Head to Head

We evaluated different RAG products and frameworks on the benchmark questions to see how they stack up. See where each one wins across the different question types.

Type 25% 50% 75% 100%
Basic -- -- -- --
Semantic -- -- -- --
Intra-Doc -- -- -- --
Project Constrained -- -- -- --
Conflicting -- -- -- --
Completeness -- -- -- --
Misc. -- -- -- --
High Level -- -- -- --
Not Found -- -- -- --

Citation

@misc{sun2026enterpriseragbench,
  title        = {EnterpriseRAG-Bench: A RAG Benchmark for Company Internal Knowledge},
  author       = {Sun, Yuhong and Rahmfeld, J. and Weaver, Chris and Desai, Roshan and Huang, Wenxi and Chen, Weijia and Butler, Mark H.},
  year         = {2026},
  howpublished = {\url{https://github.com/onyx-dot-app/EnterpriseRAG-Bench}},
  note         = {Draft. Final paper forthcoming.}
}

Think your RAG can beat the field? Run the 500-question test set against your system. Email results to joachim@onyx.app. We verify, then post you on the public leaderboard.