TurboQuant in Qdrant

Ivan Pleshkov and Jonas Schulz (Qdrant) describe TurboQuant — a rotation-based vector quantization algorithm from Google Research — shipping in Qdrant 1.18, with production-grade extensions on top of the original paper.

The Compression Ladder

Before TurboQuant, Qdrant’s quantization options:

Method	Compression	Notes
Scalar Quantization (SQ)	4×	int8 per coordinate; near-lossless
Binary Quantization (BQ)	16× or 32×	Works well on isotropic models only

TurboQuant adds a new path:

TQ Variant	Compression	Key comparison
TQ 4-bit	8×	Competitive with SQ at half the storage
TQ 2-bit	16×	Beats BQ 2-bit by 9–24 pp recall
TQ 1.5-bit	~21×	Between 2-bit and 1-bit
TQ 1-bit	32×	Beats BQ 1-bit by 9–21 pp recall

What TurboQuant Is

TurboQuant (Zandieh et al., 2026, arXiv:2504.19874) is a rotation-based quantization in the Product Quantization family:

1. Apply a random orthogonal rotation — redistributes per-coordinate variance evenly so coordinates approximate N(0,1)
2. Quantize each coordinate using a fixed Lloyd-Max codebook for the standard normal distribution — universal; no per-dataset training, no calibration set, no stored codebooks
3. Score via codebook lookup; the orthogonal rotation preserves dot products and L2 distances

Two paper variants: MSE (symmetric scoring, bit-efficient) and PROD (adds QJL projection to cancel length bias, splits the bit budget). Qdrant ships MSE — symmetric scoring is required for HNSW graph construction, and the length bias is fixed more cheaply via RaBitQ renormalization.

What Qdrant Adds

Qdrant builds a hybrid of TurboQuant and RaBitQ (Gao & Long, SIGMOD 2024, arXiv:2405.12497), the other major rotation-based algorithm:

Length Renormalization (from RaBitQ)

Vanilla MSE has a persistent length bias — quantized vectors are systematically shorter. Fix: store one per-vector scalar (ratio of original to reconstructed length), multiply back at score time. Cost: 4 bytes/vector, one multiply. Far cheaper than PROD’s full QJL projection.

Per-Coordinate Calibration (Anisotropy Compensation)

The N(0,1) rotation assumption breaks on anisotropic embeddings (real production data concentrates variance in a few directions). Fix: per-segment single pre-pass estimates a (shift, scale) pair per coordinate using the P-Square algorithm (streaming, constant memory) over a reservoir sample (Vitter’s Algorithm R). Applied as x → (x + shift) · scale to align per-coordinate distribution with the Lloyd-Max grid. Free at search time — the correction folds into a per-query precomputation; the inner scoring loop is unchanged.

Isotropic data collapses to (shift=0, scale=1) — bit-identical to vanilla TurboQuant.

L2 and Unnormalized Dot Support

Vanilla TurboQuant assumes unit-sphere inputs (cosine only). Qdrant stores the original L2 norm and reconstructs L2 via ‖q − v‖² = ‖q‖² + ‖v‖² − 2⟨q, v⟩. Cosine, dot, and L2 are first-class. L1 is not supported — random orthogonal rotation preserves L2 norm, not L1.

SIMD Acceleration

• 4-bit / 2-bit: Codebook fits in one SIMD register; lookup via pshufb. Query scalar-quantized to i8 halves for _mm_maddubs_epi16. On VNNI CPUs: VPDPBUSD; on ARMv8.2-A: SDOT
• 1-bit: RaBitQ bit-plane scoring — data packed at 128 dims/16 bytes; query transposed into B bit-planes; inner product = AND + popcount per plane, weighted by 2^b

Benchmark Results (10 Datasets, HNSW m=16 ef_construct=128)

Variant (compression)	arxiv-384	arxiv-iXL	dbp-oai	cohere	h&m	laion
f32 (1×)	0.9855	0.9419	0.9625	0.9446	0.9967	0.9897
SQ (4×)	0.9674	0.9285	0.8839	0.9014	0.9789	0.9276
TQ 4-bit (8×)	0.9442	0.9193	0.9299	0.9271	0.9739	0.9438
TQ 2-bit (16×)	0.8477	0.8227	0.8480	0.8303	0.9195	0.8349
BQ 2-bit (16×)	0.6948	0.6756	0.7332	0.6880	0.7018	0.5953
TQ 1-bit (32×)	0.7127	0.6763	0.7356	0.6300	0.8540	0.6807
BQ asymmetric (32×)	0.7070	0.5919	0.7072	0.6287	0.7802	0.5800
BQ 1-bit (32×)	0.6028	0.4683	0.6098	0.5409	0.6989	0.4762

Key patterns across all 10 datasets:

• TQ 4-bit beats SQ on 3/10 datasets (up to +4.6 pp on dbp-oai); within 2 pp on 9/10
• TQ 2-bit beats BQ 2-bit by 9–24 pp on every dataset
• TQ 1-bit beats BQ 1-bit by 9–21 pp on every dataset; still beats asymmetric BQ (8-bit query)

Practical Decision Guide

Current setup	Recommendation
SQ	Try TQ 4-bit — comparable recall, half the memory
BQ (any bit depth)	Try TQ at same storage budget — typically +10–20 pp recall
L1 similarity	Stay on SQ

Migration: config change + re-index only. One-time per-segment calibration pre-pass runs in seconds.

Related Concepts

Vector Quantization · TurboQuant · RaBitQ · BBQ · Dense Vector Retrieval · ANN · Embeddings

People

• Ivan Pleshkov — Qdrant engineer; led implementation and authored companion deep-dive
• Jonas Schulz — Qdrant co-author

Related Articles

• Elasticsearch BBQ Optimized Scalar Quantization vs TurboQuant — Elastic’s benchmark of OSQ vs TurboQuant from the Elasticsearch side