Voyage, Cohere, NVIDIA embedding models evaluated for production RAG systems

A technical team evaluated Voyage 4, Cohere Embed v4, and NVIDIA Nemotron 3 Embed models for production use. They noted that switching embedding models after deploying a large-scale index involves significant operational cost and complexity. The analysis found public benchmark scores for these models were not directly comparable, as they came from different evaluation suites. The team therefore concluded that model selection should be based on a customer's specific data and needs rather than headline scores. They treat embedding model choice as a foundational architectural decision for a retrieval system.