Embedding Atlas: Interactive Visualization for Large-Scale Embeddings

Embedding Atlas provides an interactive environment for visualizing large-scale embeddings. It allows users to explore, cross-filter, and search through embedding content and metadata within a single, unified interface.

Core Features

🏷️ Automated clustering and labeling – Understand the high-level structure of your data at a glance and navigate through clusters intuitively.

🫧 Kernel density and contours – Identify dense regions and outliers visually, eliminating the need to manually sift through raw coordinates.

🧊 Order‑independent transparency – Overlapping data points are rendered with clarity, ensuring that high-density areas remain legible rather than cluttered.

🔍 Real‑time search and nearest neighbors – Quickly locate points related to a specific query or find neighbors for any existing data point in the set.

🚀 WebGPU with WebGL 2 fallback – Leverages modern rendering APIs to maintain smooth, responsive performance even when handling millions of points.

📊 Coordinated metadata views – Interactively link and filter across different metadata columns to uncover deeper data relationships.

Explore demos and documentation at: apple.github.io/embedding-atlas

Python Installation and Usage

Install the package via pip:

pip install embedding-atlas

To launch the visualization, run:

embedding-atlas <your-dataset.parquet>

You can also integrate Embedding Atlas directly into a Jupyter notebook:

from embedding_atlas.widget import EmbeddingAtlasWidget

# Initialize the widget for a DataFrame
EmbeddingAtlasWidget(df)

NPM Installation

npm install embedding-atlas

Import the core modules:

import { EmbeddingAtlas, EmbeddingView, Table } from "embedding-atlas";

For React projects:

import { EmbeddingAtlas, EmbeddingView, Table } from "embedding-atlas/react";

For Svelte projects:

import { EmbeddingAtlas, EmbeddingView, Table } from "embedding-atlas/svelte";