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

Bo Jiang

LLM agents that retrieve external knowledge typically generate a search query as text, then run a separate embedding model to encode it into a vector. This two-model pipeline adds…