Google AI Unveils Gemini Embedding 2: A Comprehensive Multimodal Model

Google has introduced Gemini Embedding 2, a significant upgrade to its Gemini model series, which previously focused solely on text. This new model addresses the challenges of high-dimensional storage and cross-modal retrieval, essential for developers working on advanced Retrieval-Augmented Generation (RAG) systems. Gemini Embedding 2 represents a paradigm shift by allowing five different media types—Text, Image, Video, Audio, and PDF—to be represented within a unified, high-dimensional vector space. This advancement eliminates the need for separate models for each data type, streamlining the embedding process. The model supports interleaved inputs, enabling developers to combine various media types in a single request, enhancing context understanding. The input limits include 8,192 tokens for text, six images, 120 seconds of video, 80 seconds of audio, and six pages of PDF. To improve efficiency, Gemini Embedding 2 employs Matryoshka Representation Learning (MRL), which prioritizes critical semantic information in the early dimensions of the vector, allowing for effective truncation without significant accuracy loss. The model's performance has been validated against the Massive Text Embedding Benchmark (MTEB), showing improvements in retrieval accuracy and robustness across specialized domains. With an 8,192-token window, it effectively handles larger text segments, reducing context fragmentation in RAG applications.