Expand description
Clark Hash: stateless sparse-JL quantization for neural embeddings.
The library centers around ClarkHash, a deterministic codec that projects an
input vector into a low-dimensional sparse signed sketch and then applies a fixed
scalar quantizer. The resulting code can be scored asymmetrically against
floating-point queries while staying fully online and fully stateless.
The original codec and configuration names, SQuaJL and SQuaJLConfig, remain
public for compatibility with earlier experiments and papers.
See the crate-level README.md for motivation, design notes, and usage examples.
Re-exports§
pub use config::SQuaJLConfig;pub use config::SimilarityMetric;pub use error::Result;pub use error::SQuaJLError;pub use index::FlatIndex;pub use index::ScoredIndex;pub use quantized::QuantizedVector;pub use quantized::QuerySketch;pub use squajl::SQuaJL;
Modules§
- config
- Configuration types and similarity-mode selection.
- error
- Error types returned by the crate.
- index
- A simple exact-scan index over quantized vectors.
- quantized
- Encoded database vectors and prepared query sketches.
- squajl
- The core stateless codec implementation.
Type Aliases§
- Clark
Hash - Package-level name for the stateless sparse-JL codec.
- Clark
Hash Config - Package-level name for the codec configuration.
- Clark
Hash Error - Package-level name for errors returned by this crate.