When Is 0.1% Enough? Analyzing the Combined Effects of Dimensionality Reduction and Quantization on Text Embedding Compression
The study investigates combining dimensionality reduction and quantization to compress text embeddings from five models across four MTEB tasks. Results show this combination achieves near-original performance with significantly smaller storage, sometimes reducing size by 0.1% while maintaining high accuracy for classification, but the optimal compression strategy depends heavily on the specific task.