Private semantic search
Find the right model.
Keep your search to yourself.
Describe what you need. A model running in your browser creates an embedding, and the marketplace searches using encrypted similarity scores. Your readable query stays on this device.
How private search works
MiniLM produces a 384-dimensional embedding locally. The browser normalizes and quantizes it, then encrypts it using Microsoft SEAL BFV. The server embeds public catalog descriptions, multiplies encrypted query components by document components, and sums them under encryption. Your browser decrypts the resulting dot products and orders the matches.
The server receives ciphertext and evaluation keys, which allow computation but not decryption. It sees request metadata and which result you subsequently open. The query, readable embedding, decryption key, scores, and ordering are not sent back. Fresh keys are discarded after each search; the query stays in page memory until cleared or the page is closed.
Search covers up to 128 available, directly callable models. Price and context descriptions support semantic matching, but are not hard filters. Quantization makes similarity approximate. Chat and Studio providers still receive readable prompts, and FHE does not independently prove correct server computation.
Prefer explicit price, context, and latency weights? Use private routing.