krzysztof/landing_page
1
0
Fork 0
mirror of https://github.com/qdrant/landing_page.git synced 2026-10-09 21:08:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

Update qdrant-landing/content/articles/turboquant-quantization.md

Browse Source 
Co-authored-by: Abdon Pijpelink <abdon.pijpelink@qdrant.com>
This commit is contained in:
Ivan Pleshkov
2026-05-12 13:56:17 +02:00
committed by GitHub
co-authored by Abdon Pijpelink
parent 1258c41d0d
commit 18327e309e
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
qdrant-landing/content/articles/turboquant-quantization.md
+1 -1
View File
@@ -83,7 +83,7 @@ Recall, HNSW (`m=16`, `ef_construct=128`), on four representative datasets — [
TurboQuant ([Zandieh et al., 2026](https://arxiv.org/abs/2504.19874)) is a rotation-based vector quantization algorithm in the PQ family, with a clean theoretical recipe: TurboQuant ([Zandieh et al., 2026](https://arxiv.org/abs/2504.19874)) is a rotation-based vector quantization algorithm in the PQ family, with a clean theoretical recipe:
1. **Apply a random orthogonal rotation** to every vector. This redistributes per-coordinate variance evenly; after rotation each coordinate looks roughly Gaussian with the same variance. 1. **Apply a random orthogonal rotation** to every vector. This redistributes per-coordinate variance evenly; after rotation each coordinate looks roughly Gaussian with the same variance.
2. **Quantize each coordinate independently** with a fixed Lloyd-Max codebook for the standard normal distribution. One codebook of `2^b` levels for the entire dataset, hard-coded as a small lookup table. 2. **Quantize each coordinate independently** with a fixed lookup table of representative values (Lloyd-Max codebook) for the standard normal distribution. One codebook of `2^b` levels for the entire dataset, hard-coded as a small lookup table.
3. **Score** quantized vectors by reconstructing the dot product directly from the codebook indices. The rotation is orthogonal, so it preserves dot products and L2 distances — no need to ever undo it. 3. **Score** quantized vectors by reconstructing the dot product directly from the codebook indices. The rotation is orthogonal, so it preserves dot products and L2 distances — no need to ever undo it.
The elegance: **no per-dataset training, no calibration set, no codebooks to persist**. The codebook is derived once from the standard normal distribution and is universal — the same lookup table works for every dataset and every dimensionality. By contrast, PQ requires a learned codebook trained on representative data and shipped alongside the index. The elegance: **no per-dataset training, no calibration set, no codebooks to persist**. The codebook is derived once from the standard normal distribution and is universal — the same lookup table works for every dataset and every dimensionality. By contrast, PQ requires a learned codebook trained on representative data and shipped alongside the index.