Visualizing Embeddings in 2D and 3D Space
Upload JSON/CSV with pre-calculated embeddings or enter texts to embed via Ollama on-prem. View with UMAP, t-SNE, or PCA in interactive 2D and 3D scatter plots. 100% client-side upload - no data transmitted.
Data source
Projection Configuration
Algorithms and limitations
Principal Component Analysis: proiezione lineare deterministica che massimizza la varianza. Veloce, esatta, ma non cattura strutture non-lineari.
t-Distributed Stochastic Neighbor Embedding: ottimizzazione iterativa che preserva la struttura locale. Approssimazione ≤500 iter per performance browser. Non deterministico.
Simplified implementation for the browser: not identical to the original library. Adapted for visual exploration, not for production science.
Uniform Manifold Approximation and Projection: k-NN graph + force layout approssimato. Preserva struttura locale e globale. Algoritmo approx documentato - non identico a UMAP originale.
Simplified implementation for the browser: not identical to the original library. Adapted for visual exploration, not for production science.
Why use this tool?
- TF Embedding Projector () : . No Google accounts required, no uploads to the Google servers, Italian interface, all (UMAP and) t-SNE & PCA on one page.
- vs Atlas Nomic: , free with no points limit for account, on prem privacy w/ embedded Ollama.
- Local data processing, no : sent to the server (upload mode)
Come utilizzare Embedding Visualizer 2D/3D
Load your embeddings or generate from text
You have two options: load a JSON file (array of objects with "text" and "embedding") or CSV with numerical columns, or input text into the textarea and generate embeddings using Ollama on-prem (all-MiniLM-L6-v2). The texts are converted to high-dimensional vectors by the model.
Choose a Dimensionality Reduction Algorithm
Choose between PCA (linear, deterministic, fast), UMAP (non-linear, preserves global and local structure), or t-SNE (non-linear, optimal for visual clustering). Also select 2D or 3D for interactive plots.
Analyze scatter plots and clusters
Nearby points in the projection space have semantically similar text. Hover over each point to see the original text. In 3D, you can rotate and zoom the scene. Well-defined clusters indicate good semantic separation.
Suggerimenti
- For good visualization, RAG, load your knowledge base chunks along with the queries you expect to receive: separate clusters indicate that queries do not find similar responses in chunks, suggesting a need for improving chunking strategy.
- If you use t-SNE and clusters seem random, try increasing perplexity (a parameter balancing local and global structure) and letting the algorithm run for more iterations. For less than 100 points, t-SNE tends to produce unstable results.
- Use the 3D method with UMAP to explore datasets with over 1000 points: often the third dimension reveals structures and sub-clusters not visible in 2D.
Domande frequenti
What is an embed and why show it?
An embedding is a numerical representation of text (or images, audio) in a high-dimensional vector space. Texts semantically similar have nearby vectors. Visualizing them in 2D/3D helps understand how a model "sees" your data: useful for debugging RAG pipeline, evaluating the quality of embeddings and identifying unexpected outliers or clusters.
What is the difference between PCA, UMAP and t-SNE?
PCA is linear and deterministic: maintains maximum variance but loses non-linear structures. UMAP preserves local and global structure better, scalable on large datasets. t-SNE is optimal for visualizing clusters but does not preserve absolute distances and is slower. For most use cases, RAG, UMAP is the best compromise.
What format should the JSON file be in for upload?
JSON must be an array of objects with at least two keys: "text" (string containing the original text) and "embedding" (array of float numbers representing the vector). Example: [
Are my data sent to external servers?
Dimensional reduction (PCA, UMAP, t-SNE) occurs entirely in the browser (client-side) without transmitting vectors to any server. Only text embedding generation uses local backend (Ollama on-prem) for the all-MiniLM-L6-v2 model, without cloud provider.