AI tends to invent things. This is unattractive for nearly anyone who uses it repeatedly, but especially for corporations where incorrect results could harm the underside line. Half of the workers answer According to a recent Salesforce survey, they fear that the responses from their company's generative AI-powered systems are inaccurate.
Although no technique can solve these “hallucinations,” some may help. For example, Retrieval-Augmented Generation (RAG) couples an AI model with a knowledge base to offer the model with additional information before it responds, acting as a style of fact-checking mechanism.
Thanks to the big demand for more reliable AI, entire corporations have been built on RAG. Voyage AI is certainly one of them. Founded in 2023 by Stanford professor Tengyu Ma, Voyage operates RAG systems for corporations resembling Harvey, Vanta, Replit and SK Telecom.
“Voyage is committed to improving the search and retrieval accuracy and efficiency of enterprise AI,” Ma said in an interview with TechCrunch. “Travel solutions are tailored to specific areas resembling coding, finance, legal and multilingual applications, in addition to an organization’s data.”
To launch RAG systems, Voyage trains AI models to convert text, documents, PDFs and other forms of information 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 resembling RAG.
Voyage uses a special style of embedding called contextual embedding, which captures not only the semantic meaning of information, but additionally the context by which the info appears. For example, if the word “bank” appears within the sentences “I sat on the bank of the river” and “I deposited money within the bank,” Voyage’s embedding models would generate different vectors for every instance of “bank” – reflecting the difference Meanings implied by the context.
Voyage hosts and licenses its models for on-premise, private cloud or public cloud use and optimizes its models for patrons who decide to pay for this service. The company is just not unique on this regard – OpenAI also has a custom embedding service – but Ma claims that Voyage's models deliver higher performance at a lower cost.
“In RAG, when now we have an issue or request, we first retrieve relevant information from an unstructured knowledge database – identical to a librarian searches through books in a library,” he explained. “Traditional RAG methods often struggle with lack of context when encoding information, which results in errors in retrieving relevant information. Voyage’s embedding models have best-in-class retrieval accuracy, which is reflected within the end-to-end response quality of RAG systems.”
Giving weight to those daring claims is a challenge approval from OpenAI's predominant competitor Anthropic; An Anthropic support document describes Voyage’s models as “cutting-edge.”
“Voyage’s approach uses vector embeddings trained on the corporate’s data to enable contextual retrieval,” said Ma, “which significantly improves retrieval accuracy.”
Ma says Palo Alto-based Voyage has just over 250 customers. He declined to reply questions on earnings.
In September, Voyage, which employs a few dozen people, closed a $20 million Series A round led by CRV with participation from Wing VC, Conviction, Snowflake and Databricks. Ma says the money injection, which brings Voyage's total revenue to $28 million, will support the launch of latest embedding models and permit the corporate to double in size.
