Definition
Embedding models map queries and documents into dense vectors so retrieval systems can rank passages via similarity (often cosine / dot product after L2 normalization). They sit under rag, agent memory, and code search stacks.
Key Points
- Accuracy–efficiency curve: large BF16 flagships vs distilled / quantized production variants
- Context length matters for agent memory (e.g., 32k-token encoders)
- Multilingual and domain fine-tuning (legal, finance, medical, code) change real-world recall
Recent Developments
- 2026-07-16: nvidia releases nemotron-3-embed (8B BF16 #1 rteb; 1B BF16; 1B NVFP4) under OpenMDW-1.1 (nvidia-nemotron-3-embed-rteb)