Clark Hash: Stateless Sparse Johnson-Lindenstrauss Quantization for Neural Embeddings

Researchers introduced Clark Hash, a stateless codec for neural embeddings that compresses 384-dimensional vectors from 1,536 bytes to 48 bytes.

It uses sparse signed Johnson-Lindenstrauss projections plus scalar quantization, with no training, codebooks, rotations, or corpus statistics required.

On multilingual sentence-similarity benchmarks, the compressed vectors stayed highly aligned with dense cosine-similarity scores.

The approach could greatly reduce memory use and simplify deployment for similarity search systems.