Definition
RTEB (Retrieval Embedding Benchmark) evaluates embedding-model retrieval accuracy on real-world-oriented public tasks, typically reported as average NDCG@10 across 16 public tasks.
Key Points
- Designed to stress retrieval quality for production RAG and agentic search, not only academic MTEB-style suites
- Leaderboard rankings shift; scores should be dated (e.g., July 16–17, 2026 snapshots)
- Complementary benchmarks often cited alongside: mmteb, ViDoRe-V3 text, LongEmbed
July 2026 Snapshot
nemotron-3-embed flagship Nemotron-3-Embed-8B-BF16 ranked #1 at 78.46 avg NDCG@10 (2026-07-16-nvidia-nemotron-3-embed-8b-model-card).
Info
Leaderboard wins are one signal. NVIDIA and secondary coverage note real corpora, latency, and index design still dominate production outcomes.