Text similarity
Compare the meaning of two short passages with an on-device model.
On-device sentence model
Model and runtime licensesEnglish semantic matching · all-MiniLM-L6-v2 · no account or document upload.
Setup includes the local model. Checking download size… Embeddings stay in memory. Documents remain ephemeral unless you explicitly save a session.
First passage
Compare meaning
Compare short texts of up to 800 characters each. Shared vocabulary and opposite claims can still receive a high score.
The English model reads at most 256 tokens per passage. Scores measure semantic closeness and are not confidence probabilities.
Model limits and browser support
- English sentence embeddings, not a generative answer or factual-verification model. The model reads at most 256 tokens per passage.
- Up to 128 source passages / 64,000 characters and an 800-character query, or two 800-character passages. Similarity scores are not confidence probabilities.
- PDF search supports text-layer PDFs up to 25 pages / 8 MiB, requires the PDF packs and worker canvas support; scanned pages need OCR. TXT/Markdown require UTF-8.
- Physical phones and real Safari remain unverified. Desktop WASM uses one thread. Offline reload requires service-worker support; Windows WebKit offline remains unverified.
No generated answers or summaries are produced by this model. Document chat and summary have separate quality gates.