Most vector indexes (FAISS PQ, ScaNN) require a k-means training pass over your data before indexing. If the corpus shifts, you retrain and rebuild.
Google Research’s
#TurboQuant skips it. A new Rust library called turbovec just shipped a production implementation with Python bindings.
The trick: normalize each vector, apply a random rotation, scalar quantize. After rotation, coordinates follow a known
#Gaussian regardless of input. Optimal buckets become a math problem, not a data problem.
On 10M document embeddings:
• 31 GB compresses to 4 GB (16x at 2-bit)
• Recall within 0-1 point of FAISS PQ on OpenAI embeddings
• 12-20% faster than FAISS IndexPQFastScan on ARM
• Distortion within 2.7x of the Shannon lower bound
Zero training. Zero data passes. Incremental indexing.
For anyone building agentic or RAG systems: your vector layer stops forcing architectural compromises. Local hardware, shifting corpora, air-gapped deployments all work.
Repo: https://lnkd.in/eCTJ5svn
#ai
#vectors
#rag
