Last week, Cohere introduced Embed 4, an innovative artificial intelligence (AI) embedding tool designed for businesses engaged in the development and deployment of AI applications and agents. Based in Toronto, the company focuses on creating enterprise-level AI models and tools, asserting that Embed 4 is capable of comprehending complex, multimodal documents and efficiently extracting the necessary information for AI systems to execute tasks. The new tool is also touted as a means for businesses to reduce data storage expenditures by allowing the sharing of compressed embeddings instead of entire documents.
In a blog announcement, Cohere outlined the specifics of Embed 4 and officially launched the product. This multimodal embedding tool enhances the search and retrieval capabilities of existing AI systems. Businesses can access Embed 4 directly from Cohere’s website, as well as through platforms like Microsoft Azure AI Foundry and Amazon SageMaker. Furthermore, it is available for private deployment within any virtual private cloud (VPC) or on-premise settings.
Cohere’s AI models employ a technology called Retrieval-Augmented Generation (RAG) to extract information from their knowledge bases. This mechanism functions by prompting searches for specific information based on keywords, ranking methods, and other algorithmic rules. Embed 4 effectively replaces this function for data sourced from external repositories.
The company emphasizes that Embed 4 can be integrated into any pre-existing AI systems, which could include applications or agents. Typically, enterprises utilizing such tools either rely on the search engines of third-party AI models or create custom search solutions. Cohere posits that Embed 4 offers a superior alternative to both of these options.
A standout feature of Embed 4 is its support for multimodality. The tool is capable of contextually interpreting documents that include not just text, but also images, graphs, tables, diagrams, and code. Additionally, it accommodates over 100 languages, such as Arabic, Japanese, Korean, and French, allowing businesses around the world to effortlessly access their data.
Cohere also noted that Embed 4 has been trained on noisy real-world data, which encompasses imperfect documents that may contain spelling errors, formatting inconsistencies, or varying page orientations. This training enables the AI tool to retrieve such documents without compromising the accuracy of the results.
Moreover, the AI model features domain-specific knowledge relevant to regulated industries, including finance, healthcare, and manufacturing. This focus allows Embed 4 to be deployed in VPC and on-premise environments, ensuring the security of sensitive data.
