Embedding Similarity Playground

This playground demystifies text embeddings—the mathematical representation of meaning. When you input sentences, the tool doesn't just treat them as strings; it converts each one into a high-dimensional vector (a list of numbers). This process captures semantic context.

We then calculate cosine similarity, which measures the angle between these vectors. A smaller angle (closer to 1) means the sentences are highly similar in meaning, regardless of specific vocabulary. The visualization helps you see this:

• Heatmap: Shows similarity scores instantly.
• PCA Map: Projects the high dimensions onto a 2D plane for visual grouping.
• Pair Comparison: Explicitly flags your most and least similar pairs.

Crucially, all calculations happen entirely in your browser, ensuring that your input data remains fully private.

To get the most insightful results from your embedding playground session, structure your inputs thoughtfully. Think about testing specific linguistic hypotheses.

• Test Synonyms vs. Concepts: Input pairs like 'big dog' and 'large canine.' They should be highly similar, but try varying the structure (e.g., 'dog that is big').
• Contrast Related Fields: Compare a sentence from history ('The Roman Empire') with one from biology ('Mammal evolution'). The distance should be large, confirming thematic separation.
• Use Varied Lengths: Test short phrases (e.g., 'AI is cool') against long paragraphs. Observe how the model maintains core meaning despite length differences.

By intentionally creating these contrasts, you can gain a deeper appreciation for the subtle ways AI models measure semantic distance.