Voyage AI Secures $20M to Enhance Enterprise RAG with Advanced Embedding Models

Why Snowflake is backing embedding startup Voyage AI to improve enterprise RAG

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More In the world of Retrieval Augmented Generation (RAG) for enterprise AI, embedding models are critical. It is the embedding model that essentially translates different types of content into vectors, where it can be understood and used by AI and RAG approaches. OpenAI at one point dominated the embeddings space with its ada embeddings model, but some enterprises have come to realize over time that it's not specific enough for their particular use cases. That's where Voyage AI fits into the market. The startup today announced that it has raised a $20 million series A round of funding to advance the development of its embedding and retrieval models for enterprise RAG AI use cases. Among the company's backers is cloud data vendor Snowflake, which is now also set to integrate the Voyage AI models into its Cortex AI service. Specifically the Voyage AI will land in the Cortex AI search service which is based on technology from Snowflake's acquisition of AI search vendor Neeva. Voyage AI's mission is all about making enterprise RAG better. The company has a multilingual embedding model that supports 27 languages, with a high degree of accuracy. "Basically, we make RAG better by improving the retrieval quality," Tengyu Ma, founder and CEO of Voyage AI, told VentureBeat. "When you have more relevant documents, the response becomes better, because if you don't have relevant documents, then the large language model will hallucinate." How Voyage AI improves enterprise RAG with better embeddings Embedding models are nothing new and are a foundational element of large language model training and RAG deployments. Ma explained that Voyage AI is about building embedding and reranker models for improving retrieval quality. Ma argued that when it comes to RAG where specific domain or enterprise information is needed, existing approaches, particularly OpenAI's approach, isn't enough. "I think people realize that OpenAI's ada is not good enough now, because when you have higher and higher accuracy requirements, it is not accurate enough," Ma said. "So we do embeddings with better accuracy and more understanding of complex concepts." He explained that the way Voyage AI improves accuracy is with a number of advanced techniques. Voyage AI optimizes every part of the training pipeline. That includes collecting and filtering the data. Ma also noted that his company trains its models for different specific domains such as coding, finance and legal use cases. "This allows us to get even better performance for a particular domain," he said. How a contrastive learning approach improves training Training is often a particularly thorny issue as most data is unlabelled. In order to get value from unlabelled data for an enterprise, Voyage AI uses a technique called contrastive learning to train its models. Ma explained that contrastive learning is a different approach than the typical 'next word prediction' approach that is used for some training operations. In the next word approach the model predicts what word or words should follow another word or phrase based on patterns. Contrastive learning takes a different path. "You create this kind of so called contrastive pairs from unlabeled data, and use that to train the model," Ma said. Why Snowflake is embracing Voyage AI to improve enterprise RAG For Snowflake, supporting Voyage AI and integrating it into its Cortex AI services, is all about making AI more useful to enterprise users. "Every provider is trying to build some kind of a RAG system and very much the angle we take is you point us at the data, you can talk to your data, and whether it's structured or unstructured, it will just work," Vivek Raghunathan, SVP of Engineering at Snowflake told VentureBeat. Raghunathan further explained that Snowflake is excited about Voyage AI's models because of the improved and advanced capabilities that they will bring to Snowflake's customers including multilingual capabilities. He also noted that Voyage AI provides longer context windows which will also help to improve enterprise use cases. Snowflake already has its own Arctic embedding model which is currently often the default. The Voyage AI models will provide an optional alternative for users. "Think of the Pareto frontier of efficiency versus quality, our models tend to be focused for a certain size," Raghunathan said. "Voyage AI 's models are far higher quality for the really hard use cases."

Voyage AI is building RAG tools to make AI hallucinate less

AI tends to make things up. That's unappealing to just about anyone who uses it on a regular basis, but especially to businesses, for which fallacious results could hurt the bottom line. Half of workers responding to a recent survey from Salesforce say they worry answers from their company's generative AI-powered systems are inaccurate. While no technique can solve these "hallucinations," some can help. For example, retrieval-augmented generation, or RAG, pairs an AI model with a knowledge base to provide the model supplemental info before it answers, serving as a sort of fact-checking mechanism. Entire businesses have been built on RAG, thanks to the sky-high demand for more reliable AI. Voyage AI is one of these. Founded by Stanford professor Tengyu Ma in 2023, Voyage powers RAG systems for companies including Harvey, Vanta, Replit, and SK Telecom. "Voyage is on a mission to enhance search and retrieval accuracy and efficiency in enterprise AI," Ma told TechCrunch in an interview. "Voyage solutions [are] tailored to specific domains, such as coding, finance, legal, and multilingual applications, and tailored to a company's data." To spin up RAG systems, Voyage trains AI models to convert text, documents, PDFs, and other forms of data into numerical representations called vector embeddings. Embeddings capture the meaning and relationships between different data points in a compact format, making them useful for search-related applications, like RAG. Voyage uses a particular type of embedding called contextual embedding, which captures not only the semantic meaning of data but the context in which the data appears. For example, given the word "bank" in the sentences "I sat on the bank of the river" and "I deposited money in the bank," Voyage's embedding models would generate different vectors for each instance of "bank" -- reflecting the different meanings implied by the context. Voyage hosts and licenses its models for on-premises, private cloud, or public cloud use, and fine-tunes its models for clients that opt to pay for this service. The company isn't unique in that regard -- OpenAI, too, has a tailorable embedding service -- but Ma claims that Voyage's models deliver better performance at lower costs. "In RAG, given a question or query, we first retrieve relevant info from an unstructured knowledge base -- like a librarian searching books from a library," he explained. "Conventional RAG methods often struggle with context loss during information encoding, leading to failures in retrieving relevant information. Voyage's embedding models have best-in-class retrieval accuracy, which translates to the end-to-end response quality of RAG systems." Lending weight to those bold claims is an endorsement from OpenAI chief rival Anthropic; an Anthropic support doc describes Voyage's models as "state of the art." "Voyage's approach uses vector embeddings trained on the company's data to provide context-aware retrievals," Ma said, "which significantly improves retrieval accuracy." Ma says that Palo Alto-based Voyage has just over 250 customers. He declined to answer questions about revenue. In September, Voyage, which has around a dozen employees, closed a $20 million Series A round led by CRV with participation from Wing VC, Conviction, Snowflake, and Databricks. Ma says that the cash infusion, which brings Voyage's total raised to $28 million, will support the launch of new embedding models and will let the company double its size.