Visualize semantic similarity between texts
The Semantic Similarity Matrix embeds each of your texts using your own OpenAI API key, then renders an N×N heatmap of the cosine similarity between every pair. It turns a list of strings into an at-a-glance picture of which items are semantically close — invaluable for deduplication, dataset auditing, and cluster discovery.
How it works
All texts are sent in one batched request to the OpenAI Embeddings API. The returned vectors are kept in browser memory, and cosine similarity is computed locally for each pair: the dot product of two vectors divided by the product of their magnitudes. Values range from -1 to 1, where 1 means identical direction. Each cell is colour-coded — brighter green for higher similarity — so patterns jump out immediately. The diagonal is always 1.0.
What cosine similarity actually measures
Cosine similarity is not “how many words match” — it measures the angle between two high-dimensional vectors. Two short sentences about the same concept but using entirely different words can score above 0.9 with a strong embedding model; two texts sharing many surface words but on different topics may score 0.5 or lower. This is what makes embeddings useful for semantic deduplication where simple string matching fails.
Typical cosine similarity ranges for text embeddings using text-embedding-3-small:
| Score range | Typical meaning |
|---|---|
| 0.95 – 1.0 | Near-identical or paraphrased |
| 0.85 – 0.95 | Closely related, same topic |
| 0.70 – 0.85 | Related, overlapping theme |
| Below 0.70 | Distinct topics |
These thresholds vary by embedding model and domain; calibrate against known pairs from your dataset to find the right cutoff for your use case.
Practical uses
Deduplication before indexing: paste product descriptions, support articles, or job postings to spot near-duplicate entries before they reach a search index or RAG pipeline. The matrix makes it immediately visible when two chunks are functionally identical even if the wording differs.
Validating training data classes: for a classification dataset, paste a sample from each class. If classes you expect to be distinct show high cross-class similarity, the model may struggle to learn the boundary — surface that before training.
Clustering discovery: paste a diverse list of sentences to see which naturally group together. Bright clusters off the diagonal suggest the data has a hidden category structure worth making explicit.
Tips
- Look for unexpectedly bright off-diagonal cells: those are likely near-duplicates.
- If two texts you expect to be unrelated score high, your embedding model may be picking up shared domain vocabulary rather than shared meaning — check the actual texts to distinguish.
- Keep the set under 20 lines so the grid stays readable; the tool keeps all N² cells visible but heatmaps with more than ~15 rows become hard to scan.
- The embeddings request runs against OpenAI directly with your key; nothing is stored or sent to any other server.