Hasan Toor

TurboVec Beats FAISS, Packs 10M Docs in 4 GB

Vector databases are no longer a cloud product. They're becoming a pip install. A new open-source project called turbovec just crossed 10K stars on GitHub. And once you understand what it does, you understand why. It's a Rust vector index with Python bindings, built on Google Research's TurboQuant algorithm, a quantizer accepted at ICLR 2026 that compresses embeddings to within a hair of the theoretical Shannon limit. No codebook training. No train phase. No rebuilds as your corpus grows. You add vectors, they're indexed. Done. The headline number: A 10 million document corpus takes 31 GB of RAM as float32. turbovec fits it in 4 GB and searches it faster than FAISS. Read that again. Faster than FAISS. The library Meta has tuned for a decade. Hand-written NEON and AVX-512 kernels beat FAISS FastScan by 12–20% on ARM and match-or-beat it on x86. (And the recall benchmarks are published openly against FAISS as the baseline including the configs where it loses. That honesty alone is rare in this space.) But the speed isn't even the strategic part. The strategic part is what this enables: Fully local, air-gapped RAG. 10M documents in 4 GB means your entire company knowledge base fits in the RAM of a MacBook. Pair it with an open-source embedding model and nothing not a query, not a vector, not a document ever leaves your machine. It also ships drop-in replacements for the vector stores inside LangChain, LlamaIndex, and Haystack. Swap one import, keep your pipeline. The switching cost is approximately zero. The obvious comparison is SQLite. Databases used to be servers you provisioned and paid for. Then SQLite made the database a file inside your app, and an entire category of managed infrastructure became optional for most use cases. The same compression-driven collapse is now coming for vector search. Every startup selling "managed vector search" as a line item should be paying attention. When the index fits in laptop RAM, runs faster than the industry standard, and installs in one line the moat was never the database. The vector database is becoming an embedded library, not a cloud service. And the frontier of RAG just moved on-device. Really cool to see.