Snowflake Invests in Voyage AI to Optimize Multilingual RAG Applications in the AI Data Cloud

Natural language is rapidly becoming the bridge between human and machine communication. But hallucinations — when a model generates a false or misleading answer — continue to be the biggest barrier to the adoption of generative AI. Retrieval-augmented generation (RAG) allows enterprises to ground responses from LLMs in their specific organization’s data, reducing hallucinations, improving contextualized understanding and improving explainability. This approach ensures that AI-powered applications built for specific business needs deliver responses that are accurate, relevant and reliable.

In a RAG application, embeddings — representations of real-world objects, like words, images or videos, in a format computers can understand — play a crucial role in converting the user prompt into a format the model can use to capture its semantic meaning. Ultimately, this helps the model generate a response to indicate the question can’t be answered or answer the user’s inquiry based on their organization's data. 

In an era of globalization and the ongoing democratization of data insights, the ability to support conversational interactions in multiple languages is crucial. That’s why Snowflake Cortex AI, available in multiple cloud regions worldwide, is adding Voyage AI’s multilingual embedding model, which can be paired with multilingual LLMs from Meta, Mistral AI and more, to build these essential RAG applications. 

Voyage AI's multilingual embedding model, Voyage Multilingual 2, is optimized for multilingual retrieval and RAG. Supporting 27 languages, like French, German, Spanish, Korean and Japanese, it outperforms OpenAI (text-embedding-3-large); Multilingual E5 (infloat/multilingual-e5-large); and Cohere (embed-multilingual-v3.0) — surpassing the second-best model by 5.6% in retrieval accuracy, based on NDCG@10, the most widely used metric to qualify retrieval quality. In retrieval and machine learning, even small percentage gains represent significant progress. As enterprises move toward productionizing their gen AI use cases, improving this accuracy rate will help further reduce hallucinations.

Cortex AI allows users to develop apps using LLMs and embed models where the data resides to benefit from the scale, speed and data governance they've come to expect with the AI Data Cloud. With this integration, Cortex AI continues to streamline the development of multilingual AI solutions, helping global organizations across regions and clouds to tap into the power of advanced search and retrieval technologies with minimal effort. This is a game-changer for Snowflake customers building solutions designed for worldwide adoption. 

Voyage AI's vision is to revolutionize AI applications by providing fundamental building blocks — like embedding models, customization and multilingual capabilities — that enhance the performance of chatbots and AI systems. Following our investment, Snowflake customers can use Voyage AI’s highly performant models directly through Cortex functions to launch complex search queries across massive multilingual data sets and develop LLM-powered apps with unparalleled accuracy, building trust between users and AI. As Snowflake continues to deliver easy, efficient and trusted AI for the globally minded enterprise, Voyage AI will remain an integral partner.