Google unveils Gemini-based AI model for text embedding—How it works
08 Mar 2025

Google has added a new experimental "embedding" model, called Gemini Embedding, to its Gemini developer API.

The model is an innovative way to convert text inputs like words and phrases into numerical representations called embeddings.

These embeddings capture the semantic meaning of the original text and are used in various applications like document retrieval and classification, owing to their cost-effectiveness and improved latency.

A first for Google
Unique feature

While other tech giants such as Amazon and OpenAI also provide embedding models via their APIs, Google's Gemini Embedding stands out as the first embedding model trained on the Gemini family of AI models.

Google says, "Trained on the Gemini model itself, this embedding model has inherited Gemini's understanding of language and nuanced context."

This makes it suitable for a variety of use cases across domains.

Gemini Embedding outperforms previous models
Performance boost

Google claims the new Gemini Embedding model is better than its previous state-of-the-art embedding model, text-embedding-004. It also performs competitively on popular embedding benchmarks.

The Gemini Embedding can handle larger chunks of text and code at once than its predecessor and supports over 100 languages, which is double that of text-embedding-004.

It is still in experimental phase
Development stage

Despite its advanced capabilities, Google notes that the Gemini Embedding model is still in an "experimental phase" with limited capacity and subject to change.

The tech giant has stated, "[W]e're working toward a stable, generally available release in the months to come."

This indicates that further improvements and expansions are expected as Google continues to refine this innovative new tool.