# EmbeddingGemma 2 Multimodal Search & Embedding Playground (PaceBowl Studio)

> Fast, in-browser WebGPU interactive demo and developer playground for Google DeepMind's EmbeddingGemma 2 (740M parameters).

## Overview
EmbeddingGemma 2 is a lightweight, open, on-device multimodal embedding model developed by Google DeepMind. It maps text, code, images, audio, and video into a single unified 768-dimensional vector space.

## Key Technical Specifications
- Architecture: Gemma 4 based modular encoders (270M text backbone + 170M vision encoder + 300M audio encoder)
- Total Parameters: 740M
- Vector Dimension: 768 dimensions
- Matryoshka Representation Learning (MRL): Truncatable to 128d, 256d, 512d, 768d
- Context Length: 8,192 tokens
- Task Prefixes:
  - Query: `task: search_query | `
  - Document: `task: search_document | `
  - Classification: `task: classification | `
  - Clustering: `task: clustering | `
- License: Apache 2.0
- Runtime Support: WebGPU (Transformers.js v3), LiteRT, Ollama, Hugging Face Transformers, Sentence-Transformers

## Tool Features
1. Live Multimodal Search Sandbox: Test text-to-text, text-to-image, and text-to-audio cosine similarity.
2. MRL Dimension Slider: Interactively measure similarity delta across 128d, 256d, 512d, 768d and calculate vector DB RAM savings.
3. Code Generator: Instant copy-paste snippets for Python, Node.js, WebGPU, and cURL.
4. Latency & VRAM Benchmarks: Hardware footprint guide for Apple Silicon, Nvidia RTX, and WebGPU in Chrome/Edge.

Website: https://gemma.pacebowl.com/
Matrix Hub: https://pacebowl.com/