Show HN: Clark Hash, 32x smaller searchable sketches for embeddings (github.com)

Clark Hash, a new Rust package, introduces a groundbreaking way to handle neural embeddings by creating compact, searchable sketches that are 32 times smaller than traditional representations. This library implements a stateless sparse Johnson-Lindenstrauss projection combined with fixed scalar quantization, allowing for efficient storage and retrieval of embeddings. For example, a 384-dimensional f32 sentence embedding, typically requiring 1,536 bytes, can be stored as a 48-byte sketch, significantly reducing memory usage and enabling easier management of large text streams and continuous data inputs. The significance of Clark Hash lies in its ability to facilitate online semantic memory storage without the need for extensive training or recalibration of codebooks, making it ideal for applications that demand rapid processing of incoming data. Additionally, its utility in local and edge computing environments enhances its appeal, especially where bandwidth and storage limitations are crucial. With features like fast embedding integration and customizable query sketches, this package provides a versatile framework for embedding management that can be easily adapted for various machine learning workflows while maintaining high-quality performance in retrieval tasks.