What does a similarity of 0.6 mean?

Cosine similarity runs from about 0 (unrelated) to 1 (same meaning) for this model. Above 0.7 is usually the same topic, under 0.3 is usually unrelated. Compare scores with each other rather than treating any one number as a rule.

Does it work in languages other than English?

MiniLM was trained mostly on English. Other languages work a little, but scores are less reliable.

Why is the first run slower?

This tool needs a processing engine that is downloaded once and then cached: Transformers.js: model weights from the Hugging Face Hub on first use (tens of MB), then cached. Only the engine is downloaded. Your file never leaves your device.

Are my files uploaded?

No. This tool runs entirely in your browser tab. Files are read from your disk into memory and the result is generated locally.

Is it free? Are there limits or watermarks?

Free and open source (MIT). There are no file-size caps, daily quotas, accounts or watermarks.

Does it work offline?

Yes, once the page (and any engine it needs) has loaded, it keeps working without a connection thanks to the service worker cache.

Home / AI Builder / Embedding Playground

Embedding Playground

See how an AI embedding model measures meaning: a similarity grid, the closest pairs, a 2D meaning map and search, all on your device.

Free, no uploads, no sign-up. Runs 100% in your browser.

How Embedding Playground works

Everything runs in your browser. Prompts, datasets and model files are never uploaded. Each line is turned into a 384 number vector by MiniLM (all-MiniLM-L6-v2) with Transformers.js in a Web Worker. Similarity is the cosine of the angle between vectors. The map uses the two main directions in the vectors (PCA).

Input: Text. Output: CSV, JSONL.

How to use Embedding Playground

Type sentences, one per line.
Press Compare meanings (the model downloads once, about 25 MB).
Read the closest pairs, the grid and the map.
Search by meaning, then download the vectors or the grid.

Tips

Prompts, datasets and API keys never leave your device, so these tools are safe for private work.
Tokenizers and AI models download once and are cached by your browser.

Embedding Playground FAQ

What does a similarity of 0.6 mean?: Cosine similarity runs from about 0 (unrelated) to 1 (same meaning) for this model. Above 0.7 is usually the same topic, under 0.3 is usually unrelated. Compare scores with each other rather than treating any one number as a rule.
Does it work in languages other than English?: MiniLM was trained mostly on English. Other languages work a little, but scores are less reliable.
Why is the first run slower?: This tool needs a processing engine that is downloaded once and then cached: Transformers.js: model weights from the Hugging Face Hub on first use (tens of MB), then cached. Only the engine is downloaded. Your file never leaves your device.
Are my files uploaded?: No. This tool runs entirely in your browser tab. Files are read from your disk into memory and the result is generated locally.
Is it free? Are there limits or watermarks?: Free and open source (MIT). There are no file-size caps, daily quotas, accounts or watermarks.
Does it work offline?: Yes, once the page (and any engine it needs) has loaded, it keeps working without a connection thanks to the service worker cache.

Related tools

RAG Text Chunker: Split documents into overlapping chunks by characters, words or tokens for embeddings and RAG, then export JSONL, JSON or CSV.
Semantic File Search: Search many PDFs, Word files, notes and code by meaning, not just exact words, all offline.
Text Similarity & Fuzzy Match: Compare two texts with Levenshtein, Jaro-Winkler, cosine and more, or fuzzy match two lists of names.
Token Counter: Count tokens with the real GPT-4o, GPT-4, Llama 3, Mistral, Qwen, Gemma or DeepSeek tokenizer, see each token, context use and cost.