Cohere has unveiled Embed v3, their most advanced embedding model designed to transform semantic search and generative AI.
Embed v3's capacity to evaluate a document's content quality overall and ascertain how well a query fits the document's subject is one of its key features. Thanks to this feature, the model can now rank the highest-quality documents at the top, which is very useful when working with complex and noisy datasets.
Cohere's Embed v3 brings not only improved search accuracy, but also cost-efficiency. Thanks to a specially designed, compression-aware training method, businesses can efficiently handle billions of embeddings without causing a significant increase in their cloud infrastructure expenses.
Varun Kumethi, Community Champion at Cohere, quoted on X:
This special training method significantly cuts the cost of running large vector databases, making it a breeze to handle billions of embeddings without breaking the bank on cloud infrastructure.
Embed v3 offers a wide array of applications for developers. It empowers businesses to enhance search applications by improving the performance of search applications that deal with real-world, noisy data. With Embed v3, businesses can expect more accurate and relevant search results. Also, Embed v3 elevates retrievals for RAG systems. The model allows these systems to provide comprehensive and detailed responses, making them more valuable and insightful.
One of the challenges faced by generative models is their inability to connect with a company's data. Standard generative models lack knowledge about specific discussions and cannot provide meaningful summaries. Embed v3, on the other hand, bridges this gap by enabling businesses to transform their data into embeddings stored in a vector database. This enables generative models to access and retrieve relevant information, thereby delivering comprehensive and context-rich summaries.
According to Cohere, by comparing the OpenAI Ada-002 embedding model it retrieves content related to a topic, but it doesn't offer valuable information for users. On the contrary, Cohere's Embed v3 model accurately recognizes and prioritizes the most informative documents, placing them at the top of the list.
Cohere is introducing new English and multilingual Embed versions, each with distinct dimensions. These versions have set the bar high by achieving state-of-the-art performance on benchmarks like the Massive Text Embedding Benchmark MTEB and out-of-domain retrieval on BEIR.
Cohere offers the following models:
- embed-english-v3.0 (1024 dimensions)
- embed-english-light-3.0 (384 dimensions)
- embed-multilingual-v3.0 (1024 dimensions)
- embed-multilingual-light-v3.0 (384 dimensions)
These models support 100+ languages and offer seamless cross-language searching, further expanding their potential applications.
Cohere embed-multilingual-v3.0 model can be used for a variety of tasks, including embedding text in multiple languages. This repository contains the tokenizer for this model, and you can use the model either through the Cohere API, AWS SageMaker, or in private deployments.