Benchmarking Google Embeddings 2 against Open-Source Models for Multilingual Dense Retrieval and RAG Systems
A new paper benchmarks Google Embeddings 2 (GE2) against several open-source models for multilingual dense retrieval and RAG systems. GE2 achieved top performance across multiple tasks, including BEIR and an Italian RAG corpus, but exhibited significantly higher latency compared to local models. Multilingual-E5-large (mE5-L) offered comparable performance on Italian retrieval with much lower latency, making it a more practical choice for applications with strict response time requirements. AI
IMPACT Highlights trade-offs between cutting-edge performance and latency in retrieval models, guiding practical deployment choices.