Google has introduced Gemini Embedding 2, a new multimodal AI model capable of converting text, images, video, audio, and documents into a single unified embedding space. The model is now available in public preview through the Gemini API and Vertex AI platform.
Embedding models transform data into numerical vectors that represent semantic meaning. These vectors help AI systems perform tasks such as semantic search, classification, clustering, and retrieval-augmented generation (RAG).
Unlike earlier Google embedding models that focused mainly on text, Gemini Embedding 2 supports five different input types. It can process up to 8,192 tokens of text, handle multiple images, analyze short video clips, embed audio directly without transcription, and process PDF documents.
The model generates 3,072-dimensional vectors by default. Developers can also scale down the embedding size to 1,536 or 768 dimensions using Matryoshka Representation Learning, which helps balance storage efficiency and performance.
One of the biggest advantages of the new model is its ability to enable cross-modal search systems. For example, developers could search for images using text descriptions or retrieve relevant videos based on audio inputs.
According to Google, the model supports more than 100 languages and aims to improve AI systems that rely on semantic understanding across different data types.
With the launch of Gemini Embedding 2, Google continues expanding the capabilities of the Gemini ecosystem and pushing forward the development of multimodal AI infrastructure.
FAQs:
What is Gemini Embedding 2?
Gemini Embedding 2 is a multimodal AI model from Google that converts text, images, video, audio, and documents into a unified embedding vector space.
Why are embedding models important?
Embedding models help AI systems understand semantic meaning, enabling applications like semantic search, recommendation systems, and retrieval-augmented generation.
What makes Gemini Embedding 2 different?
Unlike earlier models, it supports multiple data types and enables cross-modal search by mapping different inputs into the same embedding space.
