NeMo Retriever is NVIDIA’s retrieval service that generates embeddings optimized for use with Nemotron 3 Super and other models in the Nemotron family.
NeMo Retriever handles the embedding generation component of RAG pipelines, converting documents and queries into vector representations that capture semantic meaning. The embeddings are tuned to produce high-quality matches when searching for content relevant to Nemotron 3 Super’s capabilities.
When self-hosting with Milvus, you can either call NeMo Retriever’s API to generate embeddings on-demand, or run embedding models locally for complete data privacy. Milvus then stores these embeddings and performs fast similarity search. This decoupling means you’re not locked into a single embedding service—you can switch embedding providers or fine-tune embeddings for your domain while keeping Milvus as your persistent vector storage layer.