Embedding explorer

Choose a built-in set (no download needed) or bring your own:

Your vectors: drop or open a file, or paste text.

L2-normalise scales every vector to length 1, the usual step for text embeddings: Euclidean distance then ranks neighbours exactly as cosine similarity does.

Embed sentences in the browser: type one sentence per line (optionally group | sentence) and press Embed. The page loads all-MiniLM-L6-v2 (384 dimensions, the 8-bit ONNX export, about 23 MB) with transformers.js from the Hugging Face hub the first time — the browser caches it — and embeds the lines with mean pooling and unit length, as sentence-transformers does. Sample sentences gives sixteen lines on cooking, cycling, search and the weather; they fall into four groups.

Pick a projection and 2-D or 3-D. Changing a setting re-runs it; t-SNE and UMAP animate as they optimise, and Stop halts them.

Colour by any metadata column (categories get a palette, numbers a colour ramp) or by the clusters from the Clusters tab. Search highlights labels containing the text (or metadata equal to it); Enter selects the first match. Labels shows names on the map (auto: for small sets).

Tap or click a point to list its k nearest neighbours in the original space by cosine similarity or Euclidean distance, with both numbers, and whether each is also among its k nearest on the map. Click a row to hop to that point.

In 3-D, drag to rotate (two fingers pinch to zoom on a phone, arrow keys on a keyboard) or press ⟳ to spin. ⌘/Ctrl + wheel zooms, + and − zoom, Fit resets, and a mouse drag pans a 2-D map.

How faithful is the map? — two scores, measured against the original vectors: trustworthiness (Venna & Kaski 2001, as sklearn.manifold.trustworthiness, k = 10; 1 means none of a point’s map neighbours are strangers in the original space) and the share of each point’s 10 nearest neighbours that are still among its 10 nearest on the map. Above 1,500 points they are measured on a sample.

⛶ Full screen (or F) gives the whole window to the tool — the browser’s full screen where it is allowed, otherwise a full-window layer (iPhone). The map fills the space and redraws at the screen’s resolution; Panel ◨ hides or shows the side panel (on a phone it slides up from the bottom). Tabs keep working; Esc (or ✕ Exit full screen) leaves.

Clustering runs on the full-dimensional vectors, not on the picture.

The result gives the number of clusters, the noise, inertia, the mean silhouette (Rousseeuw 1987, Euclidean; noise left out) and, against a metadata column chosen in Compare with, the adjusted Rand index (1 the same grouping, about 0 chance) and a cross-table. Colour the Explore map by cluster carries the colours over.

a − b + c, ranked by cosine with q = â − b̂ + ĉ (each vector scaled to length 1 first) and with a, b and c left out of the answers — exactly gensim’s most_similar(positive=[a, c], negative=[b]). Leaving the inputs out matters: q is very often closest to a itself, which is part of why published analogy accuracies flatter embeddings. The picture projects the five vectors onto their own two principal components: when the analogy holds, b → a and c → answer are parallel. Works on any dataset with labels — your own GloVe file or embedded sentences too.

Tested against scikit-learn 1.9.1 on the same data: PCA explained-variance ratios (to 10⁻¹⁰, all three code paths), k-means inertia and labels from fixed starts (KMeans(algorithm="elkan")), DBSCAN labels point for point, silhouette_score, adjusted_rand_score and trustworthiness; t-SNE’s per-point entropies equal log(perplexity).

Related: Eigenvectors at work (PCA by hand, the SVD and latent semantic indexing), The maths of machine learning (TF-IDF, BM25, cosine similarity, softmax and attention), Neural network designer, Statistics, Model evaluation workbench and Classic ML, step by step.

Sources: L. van der Maaten and G. Hinton, “Visualizing Data using t-SNE”, JMLR 9 (2008); L. McInnes, J. Healy, J. Melville, “UMAP: Uniform Manifold Approximation and Projection”, arXiv:1802.03426; M. Wattenberg, F. Viégas, I. Johnson, “How to Use t-SNE Effectively”, Distill (2016); J. Venna and S. Kaski, “Neighborhood preservation in nonlinear projection methods” (ICANN 2001); M. Ester et al., “A density-based algorithm for discovering clusters” (KDD 1996); T. Mikolov et al., “Linguistic Regularities in Continuous Space Word Representations” (NAACL 2013); N. Reimers and I. Gurevych, “Sentence-BERT” (EMNLP 2019).

A link opens the tool with the given settings: /tools/embedding-explorer/?<parameters>#<tab>, where the tab is explore, cluster, analogy or data. Share link (top right) builds one from what is on screen. Links carry settings and built-in dataset names — never your own vectors.

Presets for example=:

For example, https://geekychris.github.io/tools/embedding-explorer/?example=swissroll-umap opens the unrolled Swiss roll, and https://geekychris.github.io/tools/embedding-explorer/?data=words&method=pca&dim=3&color=continent#explore the word vectors in 3-D coloured by continent.