What are text embeddings?

An embedding is a list of numbers (a vector) that represents the meaning of a piece of text. The model used here, all-MiniLM-L6-v2, produces 384 numbers for any input. Texts with similar meaning produce similar vectors, even if they use completely different words. "The cat sat on the mat" and "A feline rested on the rug" end up close together in this 384-dimensional space.

What is cosine similarity?

Cosine similarity measures how closely two vectors point in the same direction, ignoring their length. It ranges from -1 (opposite meaning) through 0 (unrelated) to 1 (identical meaning). In practice, most sentence embeddings fall between 0.0 and 0.7. A score above 0.5 indicates strong semantic overlap; above 0.3 is a meaningful connection; below 0.2 is mostly noise.

What happens during encoding?

Your text is tokenized (broken into subword pieces), passed through a 6-layer transformer neural network, and the outputs are averaged (mean pooled) into a single 384-dim vector, then normalized to unit length. This all happens locally in your browser via Transformers.js — nothing is sent to a server.

Why 384 dimensions?

Each dimension doesn't have a single clean meaning (like "happiness" or "science"). Instead, meaning is distributed across all dimensions simultaneously. Concepts are represented by patterns of activation across the full vector. This is why you can't read a raw embedding — the meaning is in the geometry of the whole space, not any individual number. That's what this tool helps you explore.

The tools

Vocab Decode — Compares a vector against a set of pre-embedded reference concepts (from ~1,000 up to ~5,500 depending on the selected vocab size). Think of it as shining flashlights from different directions and seeing which ones illuminate your vector most. The top matches triangulate what region of meaning-space the vector occupies.

Iterative Search — An LLM plays a guessing game against the embedding model. It proposes candidate texts, we embed them and score how close they are to the target, then the LLM sees those scores and tries smarter guesses. Over several rounds, it converges on what the original text likely said.

Arithmetic — Vector addition and subtraction let you manipulate meaning algebraically. The famous king - man + woman ≈ queen example comes from Word2Vec, a model trained on word co-occurrence. It doesn't reliably transfer here — this is a sentence embedding model that encodes actions, context, and relationships more strongly than single-word identity. Single-word analogies often produce scattered results, but sentence-level analogies work better: try "The king ruled the country" - "The man walked home" + "The woman walked home" and you'll see the result cluster around political rule. When an analogy fails, that's informative too — it reveals which relationships this model does and doesn't encode as clean linear directions.

Differential — Instead of asking "what is this vector?", this asks "what's the difference between these two vectors?" It reveals the semantic dimensions that separate two concepts — what you gain and lose as you move from one to the other.

Projection — Squashes the 384-dim space down to 2D or 3D so you can see it. Two methods are available: PCA (instant, shows global variance directions) and UMAP (slower, reveals local cluster structure). UMAP tends to produce cleaner, more informative layouts. Your encoded vectors appear as colored dots so you can see where they land relative to the vocabulary clusters.

Reading the scores

0.6+ — very high similarity; the concepts are closely related or near-synonymous

0.4 – 0.6 — clear semantic relationship; same domain or theme

0.2 – 0.4 — weak connection; may share some context but not core meaning

< 0.2 — essentially unrelated for this model

Look for score gaps in the neighbor lists. A cluster of terms at 0.55 followed by a drop to 0.35 tells you more than the absolute numbers — the embedding strongly encodes the first cluster and only weakly relates to the rest.

Limitations

This model (all-MiniLM-L6-v2) is small and fast but has tradeoffs. It works best on English sentences of moderate length. It captures general semantic similarity well but may not distinguish fine-grained nuances (sarcasm, negation, domain-specific jargon). Larger models would do better on those, but wouldn't run efficiently in a browser.

Sentence model vs. word model: Unlike Word2Vec, which was trained on word co-occurrence and famously supports single-word analogies, this model is optimized for semantic similarity between phrases and sentences. Its 384-dim space prioritizes the meaning of full expressions over the identity of individual words. Single words like "king" produce diffuse embeddings that blend multiple senses, so classic word-level arithmetic often gives noisy results. Sentence-level operations — where there's more context for the model to work with — tend to be much more coherent.

CORS & Providers

Anthropic — works directly in the browser. No proxy needed.

OpenRouter — works directly in the browser. No proxy needed.

OpenAI — CORS blocked. You need a proxy (LiteLLM, Cloudflare Worker, etc.) or use OpenRouter instead.

Azure OpenAI — requires CORS configured on your Azure resource (Settings → Resource Management → CORS).

Local servers — must enable CORS:

llama-server --cors-allow-origin '*'
OLLAMA_ORIGINS='*' ollama serve
vllm serve --cors-allow-origin '*'
lms server start --cors