Cohere’s Embed model, Command Light model, and fine-tuning capability are now available on Amazon Bedrock, AWS's fully managed service to build generative AI applications. The platform helps simplify development, while maintaining privacy and security. Embed powers AI applications centered around enterprise search and retrieval, and is available in English and multilingual versions. Additionally, with the inclusion of Command Light and the ability to fine-tune Command models, AWS customers can scale their generative applications without the burden of managing the underlying infrastructure.

Cohere Embed: Robust Search and Retrieval

Cohere Embed is the market’s leading text representation model that improves the accuracy of search, retrieval-augmented generation (RAG), classification, and clustering.

Amazon Bedrock customers can now access the latest Embed model. What sets this model apart is its distinctive design, which evaluates how well a query aligns with a content's topic and assesses its overall quality. Traditional models often focus solely on measuring topic similarity between a query and a document. However, real-world scenarios often involve noisy data with varying content quality. For example, certain documents may offer minimal insight into topics while others provide detailed information. Unfortunately, models exclusively measuring topic similarity tend to retrieve the least informative content, resulting in a suboptimal user experience.

In contrast to other embedding models, Cohere Embed showcases superior performance, especially in challenging scenarios where noisy datasets can severely impact search and retrieval accuracy.

By integrating Embed into the Amazon Bedrock ecosystem, enterprises gain access to both generative and representative models within their secure environments. This empowers AWS customers to develop a broader range of AI applications, including text generation, summarization, classification, semantic search, retrieval for RAG systems, and more.

Command Light and Fine-tuning: Efficient Models for AI Apps

Large language models provide a valuable starting point for enterprise use cases with general generative capabilities. However, while versatile, these models may not meet the performance requirements of a focused business task and can be costly to run. AWS customers can now fine-tune both Command and Command Light models on Amazon Bedrock.

Fine-tuning is a capability to improve the accuracy of specific tasks by learning from relevant data. Fine-tuned models excel at tailoring AI solutions to specific business use cases, offering precision, accuracy, and contextual awareness that out-of-the-box models may lack. While pre-trained models suit general applications like consumer chatbots, fine-tuned models prove more performant and cost-efficient for use cases based on industry-specific needs, regulatory compliance, and nuanced domains.

Command Light is a smaller version of the Command model that is faster and cheaper to run. Fine-tuning smaller models such as Command Light facilitates scalable AI applications, enabling enterprises to meet performance requirements while benefiting from the cost advantages of smaller models. In our evaluations, fine-tuning Command Light resulted in a 60% improvement in accuracy for financial services text generation tasks. Similar positive outcomes were observed across other domains, including legal, human resources, technology, and retail.

Getting Started

To learn more and get started with Cohere’s Embed and Command Light models on Amazon Bedrock, visit the Cohere model page.