Embedding Projection Plot Calculator

Projection Results

Load data and generate a projection.

Samples0

Dimensions0

Explained variance—

Neighborhood score—

Silhouette score—

Distance preservation—

Reconstruction error—

Clusters0

Selected point—

Processing time—

Nearest Neighbors

Rank	Item	Score	Distance
Select a point or search an item.

Cluster Summary

Cluster	Size	Share	Representative
Run clustering to view a summary.

1. Embedding Data

Input format

Item column

Group column

Embedding data

Use one item per row. Numeric columns become embedding dimensions. Other columns remain metadata.

Upload CSV or JSON

Embedding preset

2. Projection Configuration

Projection method

Output dimensions

Random seed

Preprocessing

Starting component

Maximum samples

t-SNE Controls

Perplexity

Learning rate

Iterations

Early exaggeration

Iteration animation

UMAP Controls

Neighbors

Minimum distance

Metric

Epochs

Autoencoder Controls

Hidden width

Training epochs

Learning rate

Batch size

3. Clustering and Similarity

Clustering method

Clusters

Cluster using

DBSCAN epsilon

DBSCAN minimum points

Similarity metric

Nearest neighbors

Search item

Compare item A

Compare item B

Choose two items to compare.

4. Plot Customization

Color points by

Point size

Point opacity

Marker style

Point labels

Grid

Legend

Legend position

Plot background

X-axis title

Y-axis title

Z-axis title

Cluster centers

2D cluster boundaries

Responsive plot

No data loaded

Formula Used

PCA projection: Z = XW. The centered embedding matrix is multiplied by selected eigenvectors.

Cosine similarity: similarity(A,B) = (A · B) / (||A|| ||B||). Higher values indicate more similar directions.

Euclidean distance: d(A,B) = √Σ(Aᵢ − Bᵢ)². Smaller values indicate closer vectors.

Silhouette score: s = (b − a) / max(a,b). Values near one indicate clearer clustering.

How to Use

Paste embeddings or upload a supported data file. Confirm item and group column names.

Select a projection method and output dimensions. Adjust method-specific controls when needed.

Choose clustering and similarity options. Generate the plot and inspect quality metrics.

Click any plotted point to view neighbors. Export charts, coordinates, clusters, or a PDF report.

Example Data Structure

item	dim1	dim2	dim3	label
cat	0.92	0.81	0.18	animal
dog	0.88	0.79	0.20	animal
car	0.18	0.14	0.86	vehicle

Frequently Asked Questions

What is an embedding projection plot?

It converts high-dimensional vectors into two or three coordinates for visual exploration.

Which projection method should I choose?

Use PCA for global structure, t-SNE for local groups, and UMAP-style projection for balanced neighborhood views.

Why can projections look different?

Stochastic methods depend on seeds, parameters, scaling, and optimization paths.

Can the calculator process BERT embeddings?

Yes. Upload numeric BERT vectors with an item identifier and optional label column.

What does neighborhood preservation mean?

It measures how many nearby items remain neighbors after dimensionality reduction.

What does a negative silhouette score mean?

Some items may be closer to another cluster than their assigned cluster.

Why is autoencoder projection slower?

It trains a neural network before extracting the bottleneck coordinates.

Can I export an interactive plot?

Yes. The HTML export includes Plotly data and an interactive standalone chart.

Are uploaded embeddings sent to a server?

No. Calculations run in the browser unless you modify this file.