EmbeddingGemma 2 in the Browser

EmbeddingGemma 2 in the Browser

Ranks documents, compares sentences across languages and searches your own images by text, all inside your browser so nothing you type or upload leaves your device. For people exploring embeddings.

Made by Caspivs on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space runs an embedding model in your browser. It is a demo of EmbeddingGemma 2 from Google, using 4-bit ONNX weights and Transformers.js. It offers semantic search, a sentence similarity matrix and text-to-image search. Everything you type or upload stays on your device.

Who it suits: It suits developers and researchers who want to see what embeddings do without a server, and anyone who wants to search their own documents or images privately.

Industry | Who uses it | How they use it | Example
Software | Developers | Try semantic search and see the embedding model run in a page | Rank a few documents against a question
Research | Language researchers | Compare sentences across languages with the similarity matrix | Check how close a pair of translations are
Content | Editors and archivists | Rank their own images against a text query | Find pictures that match a short description
Education | Students of machine learning | Switch the embedding size and watch the effect | Compare results at 768 and 128 dimensions

You can skip it if your browser has no WebGPU, since the model needs it to run.

Where to use it:

• A recent Chrome or Edge browser: open the Space page on Hugging Face.
• Safari on iOS or macOS 26 or later: it supports WebGPU too.
• Windows, macOS or Linux: download the files and open index.html in a WebGPU browser.

Use case | Who | What they get back
Semantic search | Anyone with documents | The documents ranked against your query
Similarity matrix | Language learners and researchers | Pairwise similarity between sentences across languages
Text-to-image search | Creators and archivists | Your own images, or a sample set, ranked against a text query

How to set it up: The hosted page needs only a browser with WebGPU. Downloading a copy also needs Python and pip.

• A browser with WebGPU, such as a recent Chrome or Edge.
• Python and pip, only if you download a copy.

1. Install the huggingface_hub package with pip.
2. Download the Space's files into a folder with the hf command.
3. Move into that folder.
4. Open index.html in a WebGPU browser.

pip install -U huggingface_hub
hf download Caspivs/embeddinggemma-2-demo --repo-type space --local-dir embeddinggemma-2-demo
cd embeddinggemma-2-demo

How to use it: Open the page and wait for the model to load. Pick a tab, add your text or images, and read the ranking or the matrix. Each tab has a size selector for the embedding, with 768, 512, 256 and 128 as choices.

In the semantic search tab you also pick a task, which adds the matching instruction prefix. The choices are search, question answering, fact checking and code search.

Tool | What it does
Semantic search | Ranks documents against your query
Similarity matrix | Shows pairwise cosine similarity between sentences
Text-to-image search | Ranks images against a text query
Size selector | Truncates and renormalizes the embedding to 768, 512, 256 or 128

Example: You open the semantic search tab and add a few short documents. You type a question and choose the question answering task. The page ranks the documents by how close they sit to your question.

For newcomers and experts: If you are new, open the similarity matrix tab and type a few short sentences, some in different languages. See which pairs score as close. Then try the other tabs.

If you know the field, embeddings come from the quantized model running on ONNX Runtime Web with WebGPU. The size selector truncates the vector and renormalizes it, in the way the model card recommends. Search tasks add an instruction prefix. The quantized model's operators are only available on WebGPU, which is why a WebGPU browser is needed.

Open the page in a WebGPU browser
The 4-bit model downloads and is cached
You pick a tab and add text or images
The model embeds them on your device
Rankings or similarities are shown

Good to know: The first visit downloads a lot of data, and later visits load from the browser cache.

Title: First-time downloads
Model files: 210 MB
Image encoder: 110 MB

• Image encoder: the extra download happens on the first image search
• License: apache-2
• Cost to run: none named
• Data it reaches: nothing leaves your device, and your text and images stay in the browser
• Not for: browsers without WebGPU

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/Caspivs/embeddinggemma-2-demo
• Files: https://huggingface.co/spaces/Caspivs/embeddinggemma-2-demo/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.