Vector Embedding Visualizer

Text embeddings are dense numerical vectors that represent the semantic meaning of text. Each word or sentence is mapped to a point in high-dimensional space (e.g., 1536 dimensions for OpenAI text-embedding-3-small) where semantically similar texts cluster together. 'King' and 'monarch' have similar vectors even though they share no characters. This tool uses TF-IDF vectors as a browser-side approximation.

PCA (Principal Component Analysis) is a dimensionality reduction technique that finds the 2 directions of maximum variance in high-dimensional data and projects all points onto those axes. This preserves as much of the original structure as possible in 2D. Points that are close in the original high-dimensional space tend to be close in the PCA projection. Other methods like t-SNE and UMAP often reveal more cluster structure but require more computation.

Cosine similarity measures the angle between two vectors, ignoring their magnitude. A score of 1.0 means vectors point in the same direction (identical semantic content). A score of 0.0 means they're perpendicular (unrelated topics). Unlike Euclidean distance, cosine similarity is length-invariant — a short tweet and a long article about the same topic can have high similarity. Scores above 0.5 typically indicate meaningful semantic overlap.

TF-IDF (Term Frequency-Inverse Document Frequency) creates sparse vectors based on word frequency, weighted by how rare the word is across all documents. It's fast and runs in your browser, but it only detects shared vocabulary — it cannot understand synonyms, context, or language nuances. Neural embeddings from models like OpenAI text-embedding-3-large or Sentence Transformers capture these deeper relationships using transformer architectures trained on billions of documents.

For production use: (1) OpenAI: POST to /v1/embeddings with model='text-embedding-3-small' or 'text-embedding-3-large', (2) Anthropic: use the /v1/embeddings endpoint, (3) HuggingFace: use sentence-transformers (all-MiniLM-L6-v2 is a popular small model), (4) Google: use the text-embedding-gecko model via Vertex AI. For visualization, use UMAP or t-SNE after getting real embeddings.

1) Detecting duplicate or near-duplicate content in a document collection. 2) Exploring topic clusters in a corpus before building a classifier. 3) Debugging RAG (Retrieval-Augmented Generation) systems — understanding why certain chunks are retrieved. 4) Evaluating dataset quality — checking that training examples span the intended distribution. 5) Finding outliers in a labeled dataset.