One Model Is Enough: Native Retrieval Embeddings from LLM Agent Hidden States

Alignment Loss: Minimizes angular distance between projected vectors and teacher embeddings.
Contrastive Loss: Preserves discriminative structure across queries within a batch using InfoNCE.
Rank Distillation Loss: Transfers the teacher's document ranking preferences via KL divergence.

Enterprise AI Analysis: One Model Is Enough: Native Retrieval Embeddings from LLM Agent Hidden States