Struct EmbeddingCompressor 

pub struct EmbeddingCompressor { /* private fields */ }

Random projection matrix compressor using Achlioptas (2003) sparse projection.

Each entry of the projection matrix is +sqrt(3), 0, or -sqrt(3) with probabilities 1/6, 2/3, 1/6 respectively.

Implementations

impl EmbeddingCompressor

pub fn new(config: CompressionConfig) -> Self

Build compressor with a sparse random projection matrix (Achlioptas, 2003).

Each matrix entry is:

• +sqrt(3) with probability 1/6
• 0 with probability 2/3
• -sqrt(3) with probability 1/6

pub fn compress(&self, embedding: &[f32]) -> Result<Vec<f32>, String>

Compress a single embedding vector.

Returns an error if the input length does not match input_dim.

pub fn compress_batch( &self, embeddings: &[Vec<f32>], ) -> Result<Vec<Vec<f32>>, String>

Compress a batch of embeddings.

Returns an error if any embedding has incorrect length.

pub fn similarity_preservation_ratio( &self, a: &[f32], b: &[f32], ) -> Result<f32, String>

Approximate the similarity preservation ratio between two vectors.

Computes cosine similarity in both original and compressed spaces and returns the ratio (compressed / original). Should be close to 1.0 for high-dimensional inputs (Johnson-Lindenstrauss lemma).

pub fn config(&self) -> &CompressionConfig

Return a reference to the compression configuration.

pub fn compression_ratio(&self) -> f32

Return the compression ratio (input_dim / output_dim).