1-bit residuals shrink multi-vector indexes 13x

Using 1-bit residuals plus pruning vectors to 42%, the author compressed a multi-vector index to 1.45 GB while keeping NDCG@10 at 86.4, outperforming a larger fp16 Qwen3-Embedding-8B index on the same corpus.

2026-08-26 ~ 2026-08-26 · 2 related posts

• 1-bit residuals shrink index 13x for 1.6 point loss — tomaarsen · 2026-08-26
• Optimized LI index smaller, beats Qwen3-8B by 9 points — tomaarsen · 2026-08-26