Hierarchical Pooling Achieves 5x Lossless Compression for Late Interaction Retrieval

Recently, researcher antoinechaffin addressed the high storage and computational costs of late interaction retrieval models like ColBERT by introducing a compression method called "hierarchical pooling." This technique performs hierarchical clustering on document embeddings, iteratively merging the closest clusters to significantly reduce storage and deployment costs for large-scale retrieval systems. The approach can also be generalized to multiple modalities.

Key Details and Evolution

The hierarchical pooling technique has evolved over time. Initially, LightOn announced that introducing this method achieved 2x lossless compression for ColBERT. Later, antoinechaffin explored regularized training (or k-means pooling-aware fine-tuning) tailored for this method. As summarized by @reachsumit, this lightweight fine-tuning allows the model to achieve zero accuracy loss while drastically compressing vectors. Specifically, under 32-token compression, the model's retention rate surged from 77% without regularization to 99.4%, achieving a nearly lossless 5x compression without degrading full-token performance. The author noted that there is still significant room for further improvement.

2026-07-07 ~ 2026-07-08 · 6 related posts

• Hierarchical Pooling Achieves 5x Lossless Retrieval Compression — antoine_chaffin · 2026-07-07
• [source] Hierarchical Pooling: Compression for Late-Interaction Retrieval Models — antoine_chaffin · 2026-07-07
• [source] Lossless 5x Compression: Late Interaction Retrieval Retention Hits 99.4% — antoine_chaffin · 2026-07-07
• LightOn Achieves 2x Lossless Compression for ColBERT — antoine_chaffin · 2026-07-07
• Hierarchical Pooling Shrinks Search Index 5x, Retains 90% Quality — antoine_chaffin · 2026-07-07
• [source] Lightweight Pooling Fine-Tuning Achieves Lossless Compression for ColBERT — _reachsumit · 2026-07-08