On Friday, Google added Gemini Embedding, a new “embedded” model of text to its Gemini Developer API.
The embedding model converts text input, such as words or phrases, into numerical representations known as embeddings that capture the semantic meaning of the text. Embedments are used in a variety of applications, such as searching and classifying documents. This is because it reduces costs while improving delays.
Companies such as Amazon, Cohere, and Openai offer embedded models through their respective APIs. Although Google previously provided embedded models, Gemini Embedding is the first to be trained in the Gemini family of AI models.
“Trained in the Gemini model itself, this embedded model inherits Gemini’s understanding of language and subtle context, and is now applicable to a wide range of applications,” Google said in a blog post. “We trained models significantly more generally and delivered exceptional performance in a variety of domains, including finance, science, law, and search.”
Google claims that Gemini Embedding outperforms the performance of its previous cutting-edge embedded model, Text-dembedding-004, achieving competitive performance with its popular embedded benchmarks. Compared to Text-rebedding-004, Gemini Embedding can also accept large text and chunks of code at once, supporting twice as many languages (over 100).
Google points out that Gemini Embedding is in the “experimental stage” with limited capacity and is subject to change. “[W]The company wrote in a blog post, e’ is working towards a generally available release in the coming months.”
Source link
