/Benchmarks

Two measurements, and the method for each.

One is a public retrieval benchmark, EnterpriseRAG-Bench. The other is our own test of the thing we claim: that two people can ask the same question and each get a true answer.

01EnterpriseRAG-Bench

A public benchmark for search over company documents. Written and published by Onyx.

Louvain · 91.32%

Above every well-known system.

Louvain scores 91.32% on EnterpriseRAG-Bench.

Best well-known system, Onyx (GPT-5.4), all 500 questions
72.4%
Top of the public leaderboard, all 500 questions
93.24%

The benchmark is a synthetic company’s documents across nine sources, with questions in ten categories: Basic, Semantic, Intra-Doc, Project, Constrained, Conflicting, Completeness, Misc., High Level and Not Found.1

EnterpriseRAG-Bench · overall score, % 1
Compare with
Louvain91.32
Onyx (GPT-5.4)72.4
OpenClaw68.2
OpenAI File Search61.0
RAGFlow50.2
Amazon Q (Kendra)49.0
Azure AI Search48.4
Vertex AI Search41.9
NVIDIA AI Blueprints37.7
AnythingLLM35.6
Weaviate Verba34.5
LlamaIndex27.2
LangChain25.0
Open WebUI24.9

Measured by the benchmark’s authors

PipesHub93.24
Neurond AI92.74
codos.ai92.64
Flexor.ai (+ GPT-6.1 Sol)91.48
Louvain91.32
CDL (Causal Dynamics Lab)90.91
Unblocked88.53
AHOY Labs88.26
Skyller AI87.16
Mixedbread (+ Opus 5)86.58
Hi-Q (POSTECH)81.50

Top 10 of 42 on the public leaderboard

Above every well-known system the benchmark’s authors measured. Leaderboard read 11 November 2026

  • 500

    questions

  • 511,962

    documents

  • 9

    sources

  • 10

    question categories

  1. 1Benchmark facts and the well-known systems’ scores are as Onyx, the benchmark’s author, publishes them at onyx.app/enterpriserag-bench, read 11 November 2026: its own result with GPT-5.4 and the twelve systems in its chart. The leaderboard is the public data file at Hugging Face (last updated 3 November 2026, read 11 November 2026), ranked by overall score across its 42 systems. How answers are judged is in the benchmark’s methodology file.

02The fence benchmark

Same question, two permission sets, two different true answers. Our own test, on a copy of a real corpus.

What it measures

verify:two-readers runs on a copy of a real corpus, with some of its rooms hidden from the second reader. The test assigns the room boundaries itself. The corpus, its extraction and the statements are real.

Two people ask when an engagement ends, and are told two different dates. The extension was stated in a room the second reader cannot open, so for them the engagement still ends when they were last told it did.

Nothing leaks.Every statement and entity the restricted reader sees, the full reader sees too.
The fence does something.The restricted reader sees strictly fewer entities.
Two true answers.At least one fact holds a different current value for each reader.
Hidden means hidden.Ask about something mentioned only where the second reader can’t look, and they find nothing.

The measured figures will be published once they have been reviewed.

Monotonic. The restricted reader may never see an entity or statement the full reader cannot. A failure here would be a leak.

Strict. The restricted reader must see strictly less. Equal counts would mean the fence did nothing.

Divergent. At least one fact must hold a different current value for each reader, not merely fewer rows.

Absent. Ask about something mentioned only where the second reader can’t look, and they find nothing: no refusal that names it, no count, no kind.

03Reading

What the benchmark taught us. The write-up is on the blog.

Read the write-up on the blog

See it on your own questions.

We use Google Analytics, and Google’s and LinkedIn’s advertising tags, to see how this site is used and which ads bring people here. They set cookies only if you accept. Cookie Notice.