TurboQuant x HNSW: GloVe Scaling Curve

Live demo: https://huggingface.co/spaces/olanokhin/hnsw-turboquant-glove

This project benchmarks a TurboQuant-style vector compression pipeline on real GloVe-200 embeddings, then measures how compression changes HNSW Recall@10.

Result

Default benchmark setup:

• dataset: olanokhin/glove-6b-200d-vectors
• native dimension: 200
• encoded dimension: 256 after Hadamard padding
• corpus vectors: 25,000
• queries: 500
• HNSW: ef_construction=200, M=16

Representative result from the live demo:

Config	Compression	Recall@10 delta	Interpretation
TQ 8-bit	3.12x smaller	-0.7pp	Safe sweet spot
TQ 6-bit	4.17x smaller	-1.9pp	Aggressive sweet spot
TQ 4-bit	6.25x smaller	-5.2pp	Recall starts to break

The practical takeaway: 8-bit is the safe setting; 6-bit is the aggressive setting; 4-bit is not a drop-in replacement without reranking.

What It Does

• L2-normalizes GloVe-200 vectors for cosine search.
• Applies randomized Hadamard rotation.
• Fits Lloyd-Max scalar codebooks for 1, 2, 4, 6, 8 bits.
• Reconstructs vectors and builds a standard HNSW index.
• Plots Recall@10 vs compression ratio.

The demo intentionally exposes encoded_dim=256: native GloVe-200 must be padded to a power of two for the Hadamard transform, so real compression is below the ideal bit-width ratio.

Run Locally

python3 -m pip install -r requirements.txt
python3 app.py

By default, the app loads the prepared Hugging Face dataset:

HF_DATASET_ID=olanokhin/glove-6b-200d-vectors
HF_DATASET_SPLIT=train
HF_VECTOR_COLUMN=vector

To run against a local Stanford GloVe file instead:

GLOVE_TXT_PATH=/path/to/glove.6B.200d.txt python3 app.py

Dataset Prep

The helper script converts glove.6B.200d.txt into a Hugging Face Dataset with columns word and vector:

python3 scripts/prepare_glove_dataset.py /path/to/glove.6B.200d.txt \
  --limit 30000 \
  --push-to-hub olanokhin/glove-6b-200d-vectors

Limitations

This is a Python/Gradio benchmark, not a production vector database implementation. It reconstructs float32 vectors before HNSW. A production engine would store compressed codes directly and score them through codebook lookup with SIMD kernels.

Resume Bullet

Benchmarked TurboQuant-style HNSW compression on GloVe-200, mapping Recall@10 vs compression across 1-8 bits and identifying 8-bit as the safe sweet spot (~3.1x smaller, <1pp recall loss).