Cohere Inc., an AI startup, has unveiled Embed 4, its latest AI model designed for search and retrieval applications within assistant and agent-based AI systems.
Business operations often rely heavily on retrieval-augmented generation, a technique that enables generative AI models to fetch and incorporate the most up-to-date and accurate information in real-time. This allows large language models to respond to user queries with optimal data.
Embedding models like Embed transform data from documents into mathematical representations known as vectors, capturing dynamic, multifaceted contexts of information. In the case of Embed 4, this can include everything from text documents and images to tables, graphs, code snippets, and charts.
Cohere claims that the new model boasts an extensive context length—up to 128,000 tokens, equivalent to roughly 200 pages of text. This makes it capable of processing lengthy documents such as annual financial reports, product manuals, or intricate legal contracts. Additionally, it supports over 100 languages, including major business languages like Arabic, Japanese, Korean, and French, alongside English.
According to Cohere, Embed 4 empowers organizations to search through their unstructured documents, where much of the critical data resides. Its standout feature is the ability to generate high-quality representations of complex multimodal documents within a unified vector.
The AI startup states that Embed 4 excels in highly regulated industries like finance, healthcare, and manufacturing, showcasing domain-specific understanding. These capabilities encompass searching investor presentations, yearly financial statements, medical records, procedural diagrams, product specification documents, repair guides, and supply chain records.
Cohere also highlights that Embed 4 can handle unclear images and poorly oriented documents, adapting to noisy real-world data. The company notes that the model was trained on vast amounts of scanned documents, handwritten texts, and other damaged files. Such data types are commonly encountered during the preprocessing of multimodal data in many businesses and are typically part of manual workflows.
Agora, an AI-powered search engine serving 35,000 online stores and a customer of Cohere, leverages the model to enhance its operations. Agora says it can build better search functionalities using Embed 4’s advanced multimodal embedding capabilities.
"E-commerce data is complex, containing both images and multifaceted textual descriptions," said founder Param Jaggi. "Being able to represent our products within a unified embedding makes our search faster and our internal tools more efficient."
Embed's capabilities are crucial for accurate search and retrieval, providing power to generative AI models like Cohere's Command A, a cost-efficient model released last month. Models like Command A drive conversational assistants and AI agents but depend significantly on search engines tied to secure proprietary company data to provide relevant answers to user queries. This is essential for speeding up responses, enhancing accuracy, and reducing hallucinations.
Cohere mentions that the new Embed 4 model is now integrated into North, the company's secure AI agent productivity platform, delivering semantic search capabilities via its Compass product.
Embed 4 is available starting today on Microsoft Azure AI Foundry, Amazon SageMaker, and for private deployments.
