Gemini Embedding 2

Overview. Gemini Embedding 2 is Google's multimodal embedding model, built to turn text, images, audio, video, and documents into a single shared vector space. It's served exclusively through the Gemini API and Vertex AI as a closed-weight, proprietary model. Developers building search or retrieval systems that need to compare content across different media types in one Inference call would reach for this model rather than stitching together separate embedding models per modality.

Key strengths. The model accepts five distinct input types — text, image, audio, video, and document — in a single call and embeds them into one comparable vector space, which is unusual for embedding models that typically handle a single modality. It supports adjustable output dimensionality via Matryoshka Representation Learning, defaulting to 3,072 dimensions but reducible for cases that prioritize storage footprint or lookup speed over maximum fidelity. The 8,192-token input limit is generous enough for typical passages, product descriptions, or multi-page document chunks.

When to use it. Use this model when building semantic search, recommendation, or retrieval-augmented generation pipelines that need to compare mixed-media content — for example, matching a text query against a corpus containing PDFs, images, or audio clips — inside one embedding space. It's a good fit on Google Cloud stacks already using Vertex AI, and the adjustable output dimensionality makes it suitable when you need to trade vector size against retrieval precision for a large index. It is not a generative model, so it shouldn't be reached for when the task requires producing text, code, or structured completions.

Limitations. The 8,192-token input cap means very long documents need chunking before embedding. Google has not published a parameter count, training data cutoff, or architecture details for this model, and it's available only as a hosted API on Gemini API/Vertex AI with no self-hosting or open-weight option. As an embedding model, it produces only numeric vectors — it has no streaming, tool-calling, or structured-output behavior.

Sovereignty note. Gemini Embedding 2 is created by Google and served through Google's own Gemini API and Vertex AI infrastructure. The research findings available don't specify data residency options or CLOUD Act exposure for this model, so no sovereignty claim beyond creator attribution can be made here.