Skip to content

Releases: dharmeshdev1/unsaid

Unsaid 0.1.0

Choose a tag to compare

@github-actions github-actions released this 01 Nov 16:27

The first release. Unsaid is a search engine that cannot read your queries: your question is encrypted on your own device with lattice cryptography, and the server finds relevant documents and returns them without learning the question or the results.

Try it without installing anything: https://dharmeshdev1.github.io/unsaid/

Or download a binary below and run:

unsaid demo --watch

What it does on the BEIR benchmark: 90 to 98% of exact search quality with the same embedding model, including 95.8% over 2.68 million passages. Every number comes from a command in the repository that you can run yourself.

Please read the security notes before relying on this. It has not had an independent audit. It hides your question and results from the server, but not your IP address or when you search.

Each archive has a .sha256 file beside it. The x86-64 builds need a CPU with AVX2, which means roughly 2013 or newer.

Embedding models

Choose a tag to compare

@dharmeshdev1 dharmeshdev1 released this 01 Nov 16:26

Packed embedding models, which unsaid downloads on first use and checks against hashes compiled into it. You do not need to fetch these by hand.

File Source License SHA-256
all-MiniLM-L6-v2.usm sentence-transformers/all-MiniLM-L6-v2 Apache 2.0 b3fd40c159f2dca22fdeba20ed584e64ed980208be96486068f3789d8c32b51f
bge-small-en-v1.5.usm BAAI/bge-small-en-v1.5 MIT eefe4ec54fc8f2d16dff877213276a081f64de9b44c3931bf4b1139894d0395d

Weights are stored as 8-bit integers with one scale per row. crates/unsaid-embed/tests/reference.rs checks the results against the original models.