September 2024 Summaries

Voyage has introduced the Voyage 2 series of rerankers, rerank-2 and rerank-2-lite, designed to enhance the accuracy and efficiency of retrieval systems by refining search result rankings. These rerankers, when combined with OpenAI's v3 large embedding model, increase accuracy by 13.89% and 11.86%, respectively, significantly surpassing the improvements offered by the latest Cohere reranker. They offer extended context lengths of 16K and 8K tokens, enabling better handling of query-document interactions. Rerank-2 is optimized for quality, while rerank-2-lite focuses on reducing latency while maintaining performance, both supporting multilingual capabilities across 31 languages. Evaluated over 93 domain-specific datasets, these models consistently outperform competitors in various domains, including technical documentation, finance, law, and multilingual contexts. They offer flexible, token-based pricing, and users of previous Voyage rerankers can upgrade to these new models to benefit from improved performance and enhanced context length without additional costs.

Voyage has introduced its latest embedding models, voyage-3 and voyage-3-lite, which significantly enhance retrieval quality, latency, and cost-efficiency compared to existing models like OpenAI v3 large. The voyage-3 model outperforms OpenAI v3 large by 7.55% across various domains such as code, law, finance, multilingual, and long-context, while reducing costs by 2.2 times and embedding dimensions by three times, leading to lower vectorDB costs. Additionally, voyage-3-lite offers 3.82% better retrieval accuracy than OpenAI v3 large at a fraction of the cost and supports a 32K-token context length, four times that of OpenAI. Both models are part of the Voyage 3 series, which follows the Voyage 2 series known for its domain-specific models, and aim to provide superior performance and affordability in data retrieval tasks. They have been rigorously evaluated across 40 domain-specific datasets and 26 languages, demonstrating particularly strong performance in multilingual contexts while maintaining lower costs and latency compared to competitors.