Embedding & Retrieval
Embedding models turn text (or other data) into vectors that can be compared for semantic similarity — the foundational layer beneath every RAG (retrieval-augmented generation) pipeline, semantic search product, and recommendation system. This category is judged on retrieval-benchmark performance (MTEB and similar), embedding dimensionality/cost tradeoffs, and how well a general-purpose embedding model transfers to a specific domain (code, legal, medical) without fine-tuning. GROUNDING tracks new embedding-model releases and their retrieval-benchmark results, since a weak embedding model silently caps the quality of every RAG system built on top of it, regardless of how good the generation model is.
At a glance
- 11 tracked models
- Most recently updated: Qwen3-Embedding 4B (2026-07-08)
Most actively covered
- BGE-M3
- BGE
- LaBSE
- Qwen3-Embedding-8B
- Qwen3-Embedding
- multilingual E5
- BGE-large
- E5
- NV-Embed
- Qwen3-Embedding 4B
FAQ
What is the Embedding & Retrieval category?
Models that produce vector embeddings for semantic search and RAG.
How many models are in this category?
GROUNDING currently tracks 11 models in Embedding & Retrieval.
How current is this hub?
The most recently updated entry is Qwen3-Embedding 4B, last mentioned 2026-07-08; the radar refreshes hourly.