trusty-embedder

Shared text-embedding abstraction for the trusty-* family of projects.

Provides a single async Embedder trait and a production-ready FastEmbedder implementation backed by fastembed-rs (all-MiniLM-L6-v2, 384-dimensional output, INT8-quantized ONNX) with LRU caching and ORT warmup. A MockEmbedder test double is available behind the test-support feature.

Installation

cargo add trusty-embedder

For tests/benches that want a deterministic stand-in:

cargo add trusty-embedder --features test-support

Quick Example

use trusty_embedder::{Embedder, FastEmbedder};

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let embedder = FastEmbedder::new()?;
    let vectors = embedder
        .embed_batch(&["hello world".into(), "trusty".into()])
        .await?;
    assert_eq!(vectors[0].len(), trusty_embedder::EMBEDDING_DIM);
    Ok(())
}

Feature Flags

test-support (optional) — exposes MockEmbedder, a deterministic hash-based embedder for unit tests that don't want to load an ONNX model.

Notes

Model files are downloaded by fastembed on first use and cached in the standard fastembed cache location.
The crate falls back from AllMiniLML6V2Q (INT8) to AllMiniLML6V2 (FP32) when the quantized variant is unavailable in the host fastembed build.
embed_batch is the single primitive — single-text embedding is a thin helper that wraps a one-element batch.

License

Licensed under the Elastic License 2.0.

Repository

https://github.com/bobmatnyc/trusty-common