Voyage AI Inc., a startup with software for organizing the data processed by artificial intelligence models, today announced that it has raised $20 million in funding.
CRV led the Series A investment. It was joined by Snowflake Inc., Databricks Inc., Wing VC, Conviction, Pear VC, Tectonic Capital, Mayfield Fund and Fusion Fund. Voyage AI’s total outside funding now stands at $28 million.
Before records such as business documents can be processed by an AI model, they have to be turned into so-called embeddings. Those are specialized files that represent data in a form that neural networks can understand. Under the hood, an embedding is a collection of numbers that can be upwards of several megabytes in size.
Embeddings highlight relationships between the data points they store. If an AI ingests a snippet of text that contains references to iPhones and iPads, embeddings can indicate that both devices are made by the same company. Such pointers make it easier for neural networks to reason about the data.
The process of turning files into embeddings is done not manually but rather using specialized AI models. Palo Alto, California-based Voyage AI sells a half dozen such models that customers can deploy in the cloud or on-premises. The company also offers the ability to customize its embedding generators for an organization’s specific requirements.
Alongside its funding announcement today, Voyage AI launched two additions to its AI model portfolio: voyage-3 and voyage-3-lite. The company says that both outperform OpenAI’s flagship algorithm for generating embeddings.
One of the areas where Voyage AI promises to provide better performance is retrieval quality. When an AI model receives a question from a user, it searches its knowledge repository for embeddings that contain relevant data and uses this data to generate a response. Retrieval relevance is a metric that measures the effectiveness with which an AI model can find relevant data in its embeddings.
Voyage AI says that voyage-3, the first AI model it debuted today, provides 7.55% better retrieval quality than OpenAI’s competing offering. Furthermore, the new algorithm costs 2.2 times less to use.
According to Voyage AI, voyage-3 can also reduce customers’ data storage expenses by generating embeddings with lower dimensionality than OpenAI’s offering. Dimensionality is a term for the memory footprint of embeddings. The smaller the embedding, the less it costs to store.
The other embedding generation model that Voyage AI debuted today, voyage-3-lite, is a more affordable version of voyage-3. It generates models with significantly lower dimensionality to cut users’ storage costs. This increased cost-efficiency comes at the expense of retrieval quality, but Voyage AI says that voyage-3-lite still outperforms OpenAI’s flagship embedding model.
Alongside embedding generators, Voyage sells so-called rerankers. Those are AI models that search engines and certain other applications use to improve the relevance of their results. After a search engine retrieves a set of webpages or files in response to a user query, it uses a reranker to move the most relevant items to the top of the results page.
Voyage AI debuted two new rerankers today alongside its latest embedding models. The first organizes search results with 7.14% better relevance than a comparable model from Cohere Inc. The other new reranker, in turn, provides 5.12% to 11.86% better performance than the Cohere algorithm.
Voyage AI will use its newly raised funding to expand its product portfolio with more AI models.