Turbovec: a Rust vector index that fits 10M embeddings in 4GB of RAM
The open-source index hits FAISS-beating speed on Apple silicon near 0.997 recall and drops into LangChain — but it's a days-old project with thin docs.
Copy markdown
31GB of vectors, now 4GB of RAM
Turbovec stores a 10M-document corpus in 4GB where float32 needs 31GB — an ~8x cut, built on Google's TurboQuant, a data-oblivious quantizer with no training step or index tuning to babysit.
It beats FAISS on Apple silicon
On an M3 Max it runs 10-19% faster than FAISS IndexPQFastScan across configs and edges it on accuracy — both reach ≥0.997 recall by k=4. On x86 it wins 4-bit by ~5% and trails slightly at 2-bit.
16x smaller per vector
A 1536-dim OpenAI embedding drops from 6,144 bytes to 384 at 2-bit — 16x. That's the line between a RAG index that lives in RAM on one box and one that needs a cluster or a paid vector DB.
pip install, then wire it to LangChain
`pip install turbovec` (or `cargo add turbovec`), MIT-licensed, with LangChain, LlamaIndex, Haystack and Agno adapters plus query-time filtered search — enough to stand up a local, air-gapped RAG this weekend.
The catch: it's days old and solo
It's a one-author project that just hit the Hacker News front page; commenters flag thin docs and missing SQLite/DuckDB/LanceDB bindings. Great for a side build — pin the version and benchmark before it touches prod.