I've been experimenting with an approach to vector indexing where the HNSW graph nodes store quantized embeddings (~388 bytes each at dim=1024) instead of float32 vectors (~4,096 bytes). The key insight: if you quantize embeddings using Lloyd-Max s…
Most embedding models are not Matryoshka-trained, so naive dimension truncation tends to destroy them. I tested a simple alternative: fit PCA once on a sample of embeddings, rotate vectors into the PCA basis, and then truncate. The idea is that PCA con…
We just shipped an autotune CLI for turboquant-pro — it connects to your PostgreSQL database, samples embeddings, sweeps 12 compression configurations, and tells you exactly which one to use. bash turboquant-pro autotune \ –source "dbname=mydb us…
TL;DR: We built an open-source toolkit that compresses high-dimensional vectors (embeddings, KV cache, anything in pgvector/FAISS) by 5-42x while maintaining 0.95+ cosine similarity. Benchmarked 6 methods on 2.4M real embeddings. MIT licensed. GitHub: …