"Voyage AI Enhances Snowflake’s Cortex AI with Multilingual Embeddings for Superior Enterprise RAG"

In the realm of Retrieval Augmented Generation (RAG) for enterprise AI, embedding models play a pivotal role.

These models convert various content types into vectors, making them interpretable by AI systems. While OpenAI's ada embedding model was a frontrunner in this space, many enterprises have discovered that it lacks the specificity needed for certain applications. This is where Voyage AI comes into play.

Recently, Voyage AI announced a $20 million Series A funding round to enhance its embedding and retrieval models tailored for enterprise RAG applications. Significant backing comes from Snowflake, which plans to integrate Voyage AI's models into its Cortex AI service. This integration will enhance the Cortex AI search functionality, leveraging technology from Snowflake’s acquisition of AI search company Neeva.

Voyage AI is dedicated to enhancing enterprise RAG capabilities. Its multilingual embedding model supports 27 languages with remarkable accuracy. “We improve RAG by enhancing retrieval quality,” said Tengyu Ma, founder and CEO of Voyage AI. “More relevant documents lead to better responses; without them, large language models may generate inaccurate outputs.”

Enhancing Enterprise RAG with Superior Embeddings

Embedding models are essential for training large language models (LLMs) and implementing RAG systems. Ma emphasized that Voyage AI focuses on creating advanced embedding and reranking models to improve retrieval quality for domain-specific information. He pointed out that as enterprise accuracy demands intensify, existing solutions, including OpenAI’s ada, fall short. “Our embeddings offer greater accuracy and deeper understanding of complex concepts,” Ma explained.

Voyage AI enhances accuracy through advanced techniques, optimizing the entire training pipeline, including meticulous data collection and filtering. The company tailors its models for specific industries such as finance, coding, and law, achieving superior performance in these areas.

The Role of Contrastive Learning in Training

Training machine learning models can be challenging, especially with unlabelled data. To leverage this type of data effectively, Voyage AI employs contrastive learning, a technique distinct from traditional next-word prediction methods. “We create contrastive pairs from unlabelled data to train our models,” Ma shared.

Snowflake’s Partnership with Voyage AI

For Snowflake, partnering with Voyage AI and incorporating its models into Cortex AI services is about enhancing utility for enterprise users. "Every provider is striving to develop RAG systems. Our approach enables users to interact seamlessly with their data, whether structured or unstructured," stated Vivek Raghunathan, SVP of Engineering at Snowflake.

Raghunathan expressed excitement for Voyage AI’s models due to their advanced capabilities, including multilingual support and extended context windows, which are crucial for enterprise applications. While Snowflake offers its Arctic embedding model, Voyage AI provides a compelling alternative for users. "Consider the balance between efficiency and quality; our models excel at addressing challenging use cases," Raghunathan concluded.