Tackling AI's Tall Tales: Voyage AI's New Approach

Founded in 2023 by Stanford professor Tengyu Ma, Voyage AI aims to tackle one of the major cons of AI: it's tendency to produce inaccurate information, known as "hallucinations". These are a particular concern for businesses with their reputation and profit on the line. A recent Salesforce survey found that half of employees are concerned about the accuracy of their company's AI systems.

‍

To address this problem, Voyage AI focuses on a method called retrieval-augmented generation (RAG). This technique pairs AI models with knowledge bases to provide extra information before generating answers, acting like a fact-checking tool to improve reliability.

‍

Voyage AI enhances this method by training their AI models to turn various types of data, including text, PDFs, and more, into numerical representations called vector embeddings. These help capture the meaning and relationships between different data points. They use a specific kind of embedding called contextual embedding, which takes into account not just the literal meaning but also the context in which data appears.

‍

The company offers their models for use on-site, in private clouds, and in public clouds, with options for customisation. While this isn't unheard of – OpenAI having already offered a tailorable embedding service prior to this – Ma claims that Voyage's models perform better and are more cost-effective than those of competitors. Their embedding models are said to have top-notch retrieval accuracy, leading to better overall responses in RAG systems.

‍

Voyage AI has gained significant traction as a startup, gaining support from Anthropic, a competitor to OpenAI and raising $20 million in Series A funding, bringing their total funding to $28 million. This investment will help launch new embedding models and enable the company to expand their team, which currently has only around a dozen people onboard.