Get the model

The library never downloads anything. It reads the model from a folder you give Embedder::load, and embed::catalog() tells you what that folder must hold: the six files of EmbeddingGemma-300m in ONNX, q4 quantised (about 219 MB), from onnx-community/embeddinggemma-300m-ONNX on Hugging Face, each pinned to commit 5090578d9565bb06545b4552f76e6bc2c93e4a66 with its sha256 and size.

The easy way: the search example

The search example downloads the model the first time it runs, checks every file’s sha256 and keeps it for next time:

cargo run --release --example search -- a hidden area

It keeps the model in $XDG_DATA_HOME/pito-embed/embeddinggemma-300m-q4 (~/.local/share/… when that isn’t set) on Linux and ~/Library/Application Support/pito-embed/embeddinggemma-300m-q4 on macOS. Set PITO_EMBED_MODELS to use another folder:

PITO_EMBED_MODELS=models/embeddinggemma-300m-q4 cargo run --release --example search -- a hidden area

Pass that folder to Embedder::load.

In your own program

Each catalog entry is a fetch::FetchSpec from pito-fetch: the file’s url, its sha256, its size_hint in bytes and its dest_rel, a path that starts with models/embeddinggemma-300m-q4/. Strip that prefix to get where the file goes inside the model folder. examples/search.rs does it with pito-fetch, which resumes a broken download and checks the hash. Add it next to embed:

fetch = { git = "https://github.com/gmrdad82/pito-fetch.git", tag = "v0.1.2" }

Then:

use fetch::{CancelToken, Fetcher, Request};

let fetcher = Fetcher::new(&dir);
let cancel = CancelToken::new();
for spec in embed::catalog() {
    let rel = spec.dest_rel.trim_start_matches("models/embeddinggemma-300m-q4/");
    let request = Request {
        url: spec.url.to_string(),
        dest_rel: rel.to_string(),
        sha256: spec.sha256.to_string(),
        size_hint: spec.size_hint,
    };
    fetcher.fetch_request(&request, &cancel, |_progress| {})?;
}

Any downloader works, as long as it checks the sha256. The folder’s layout is in Formats.

The model’s terms

EmbeddingGemma is Google’s model, provided under and subject to the Gemma Terms of Use and the Gemma Prohibited Use Policy. pito-embed ships no weights; whoever downloads and uses the model accepts those terms, which the MIT licence of the code doesn’t cover.