this_file: README.md

Uubed Rust workspace

One native implementation for compact byte encoding and translation-memory search.

The extension is imported as uubed_native. Model inference, model identity, TMX streaming and atomic index publication live in the companion uubed Python package. There is no second Candle/Qwen model runtime or fixed-1024 database.

cargo test --workspace
cargo clippy --workspace --all-targets --features simd -- -D warnings
uvx maturin build --release

Requires Rust 1.88+; the existing core remains edition 2021. Python integration is tested on 3.12. Native binding behavior is tested through the built wheel, rather than linking a Rust executable against a Python extension module.

SQLite stores BINARY-collated source strings, all target variants and their path/unit provenance. One int8 BLOB per English source has its positive scale and norm stored alongside it. Metadata records the format and full embedding-space identity. Query vectors must match the dimensions and contain finite, nonzero values. Search ranks cosine similarities after filtering by target language. Scale cancels in cosine, so no float32 corpus matrix is materialized. Search is exhaustive O(Nd), not an approximate-nearest-neighbor service.

See the workspace guide for the complete two-consumer workflow.

Releases and local data

./publish.sh --dry-run verifies the next release without pushing or uploading. ./publish.sh commits, tags and publishes it. See RELEASING.md for credentials, same-tag retries, dependency order and private-data exclusions.