Embeddings

The Embeddings page lets you generate text embeddings and visually compare how semantically similar two pieces of text are.

Sidebar item: Embeddings Route: /embeddings

What Are Embeddings?

An embedding is a list of numbers (a vector) that represents the meaning of a piece of text. Texts with similar meanings produce vectors that point in similar directions. LLMxRay uses this concept to let you explore semantic similarity hands-on.

Embedding Text

1. Select an embedding model from the dropdown. Only models with embedding capabilities appear (e.g., nomic-embed-text, all-minilm).
2. Enter text in the input area.
3. Click Embed. The resulting vector is displayed as a visual bar chart.

No embedding model?

If no models appear in the dropdown, you need to pull one:

ollama pull nomic-embed-text

Comparing Two Texts

1. Enter text in both input areas.
2. Click Compare. Both texts are embedded and compared.
3. The cosine similarity meter shows how semantically close the two texts are:
   • 1.0 — Identical meaning
   • 0.7+ — Very similar
   • 0.3–0.7 — Somewhat related
   • < 0.3 — Unrelated

Embedding Images

Since Ollama 0.40, /api/embed accepts an image alongside the text, on models that report both the embedding and vision capabilities. When the selected model does, an Attach an image control appears under the playground text box and under each side of the Similarity Calculator. A text-only model shows no image controls at all, because the daemon answers media with a 501.

What you can do with it:

• Embed a photo and a sentence that describes it, and compare them. Same vector space, so the cosine similarity is meaningful.
• Compare two photos.
• Embed an image alone, with no text.

Every result shows the prompt tokens it cost. An image is a fixed block of tokens on top of the text — measured on embeddinggemma-2: a short sentence cost 7 tokens, the same sentence with a tiny image cost 265. That block is paid whatever the image size.

The only model that does this today, embeddinggemma-2, runs on Ollama's MLX runtime, which exists on Apple Silicon only. On Windows or Linux ollama pull embeddinggemma-2 fails with "this model requires MLX support". To use it from another machine, forward the Mac's daemon over SSH and point LLMxRay at it in Settings:

ssh -L 11435:localhost:11434 user@your-mac

then set the Ollama URL to http://localhost:11435/api.

Use Cases

• Understanding embeddings — See what these numbers actually represent
• Testing RAG relevance — Check if your documents will match user queries
• Exploring synonyms — See how the model understands semantic relationships
• Educational — Great for AI/ML coursework on vector representations

Tips

• Different embedding models produce different vector dimensions and similarity scores.
• Short, focused texts produce more meaningful comparisons than long paragraphs.
• Results are stored in the session and can be reviewed later.