Definition
While any network layer can produce embeddings, embedding models are trained specifically so similar meanings land nearby in vector space — often with contrastive learning.
They power semantic search, RAG retrieval, deduplication, and clustering.
In simple terms
A mapmaker drawing cities so travel time matches distance on the page — proximity means relatedness.
Where you see it
- OpenAI text-embedding models and open sentence-transformers.
- Multilingual embedders for Somali–English retrieval.
How it works
1.Encode inputs
Pass text through the encoder.
2.Pool to a vector
Mean pooling or CLS embedding.
3.Compare with similarity
Cosine similarity or dot product for ranking.
Why it matters
- RAG and semantic search quality depend heavily on the embedding model you choose.
Often confused
Any LLM hidden state is an equally good retriever embedding.
Specialized embedding models usually outperform raw LLM states for retrieval.