– Embeddings are numerical representations of real-world data that help foundation models understand relationships within data
– Text embedding models are essential for natural-language processing applications and power various Google Cloud services
– Google Cloud has introduced new Vertex AI text embedding models with improved performance, dynamic embedding sizes, and competitive pricing; customization support for these models is coming soon.
Embeddings are numerical representations of real-world data, such as text, speech, image, or videos, which help foundation models powering generative AI understand relationships within data. These fixed-dimensional vectors demonstrate geometric distances between two vectors in a vector space, reflecting the relationships between real-world objects they represent. Text embedding models are crucial for various natural-language processing (NLP) applications, including document retrieval, similarity measurement, classification, and clustering, and are utilized across Google Cloud services like BigQuery and Vertex AI Search.
At Google Cloud Next ’24, two new Vertex AI text embedding models have been introduced for public preview, showcasing improved performance across different tasks. The English-language embedding model showed enhanced performance on the MTEB benchmarks, surpassing existing entries with an average score of 66.31%. Similarly, the multilingual embedding model demonstrated improved performance on the MIRACL benchmarks, achieving an average score of 56.2%. These models come with a pricing of $0.000025/1,000 characters for online requests and $0.00002/1,000 characters for batch requests.
The new text embedding models offer dynamic embedding sizes, allowing users to choose smaller dimensions with minor performance loss to save on computing and storage costs. Additionally, customization support for stable model versions is also available, with significant quality gains shown through parameter-efficient tuning methods. Users are encouraged to try out the latest models through Google’s public documentation and provide feedback for further improvements.