Unblocked
Azure AI Search
Vertex AI Search
LlamaIndex
LangChainBuild it yourself?
We tested Unblocked on 500 questions from EnterpriseRAG-Bench. Each question requires company knowledge from docs, conversations, tickets, or meetings, not just the code.
| System | Overall | Correctness | Completeness | Recall | Extra docs |
|---|---|---|---|---|---|
| Unblocked | 88.5 | 91.2 | 93.2 | 94.7 | 0.5 |
| Azure AI Search | 48.4 | 56.4 | 57.6 | 64.3 | 3.3 |
| Vertex AI Search | 41.9 | 49.2 | 55.5 | 61.8 | 4.0 |
| LlamaIndex | 27.2 | 32.4 | 37.8 | 30.6 | 1.5 |
| LangChain | 25.0 | 31.0 | 35.6 | 36.4 | 3.1 |
Overall = correctness × completeness, averaged per question. Extra docs = irrelevant documents cited per answer, lower is better.
Inside the dataset
500 questions in 10 categories over 511,962 documents from nine enterprise tools. Off-topic threads, half-finished drafts and near-duplicate pages are mixed in on purpose, the clutter a real retrieval system has to ignore.
Documents per source, and how often Unblocked cited the gold documents from each.
| Source | Documents | Our recall |
|---|---|---|
| Slack | ~275K | 88.9% |
| Gmail | ~120K | 91.7% |
| Linear | ~35K | 96.8% |
| Google Drive | ~25K | 90.6% |
| HubSpot | ~15K | 97.1% |
| Fireflies | ~10K | 80.4% |
| GitHub | ~8K | 91.7% |
| Jira | ~6K | 91.7% |
| Confluence | ~5K | 90.8% |
Methodology
- How the data was built
- Onyx scaffolded a synthetic AI-inference company, generated documents across nine tools with shared projects, people and codenames, then added noise: misfiled files, stale facts and near-duplicates with conflicting details.
- How it’s scored
- LLM judges grade each answer for correctness and completeness against the gold answer.
Benchmark: Sun et al., “EnterpriseRAG-Bench: A RAG Benchmark for Company Internal Knowledge,” arXiv:2605.05253, MIT-licensed. All scores, including Unblocked’s, are from Onyx’s published leaderboard, all 500 questions, .