Overview
Production retrieval stack evolution in 2026: open embedding collections, multilingual leaderboards (rteb, mmteb), and agentic memory use cases.
Timeline
- 2026-07-16: nvidia nemotron-3-embed tops rteb with open 8B/1B/NVFP4 family (nvidia-nemotron-3-embed-rteb)
Key Players
- nvidia
- Hugging Face leaderboard ecosystem
- Enterprise RAG / agent memory vendors evaluating NIM paths
Analysis
Leaderboard wins accelerate trial adoption, but NVIDIA’s own caveats apply: corpus fit, latency SLOs, and tiered index design (1B recall + 8B hard queries) matter more than a single NDCG number.