#88th of 12 in Vector databases
turbovec
In-process quantized vector index for Python and Rust using TurboQuant
- Stars
- 17.4k
- License
- MIT
- Last commit
- Oct 2026
- Language
- Rust
Missed the latest GitHub check; these facts may be older.
Overview
turbovec is a library that stores embeddings at 2 or 4 bits per dimension using Google Research's TurboQuant and searches the compressed form with SIMD kernels. It needs no training step, supports single-vector add and remove, and persists incrementally with sync(). It does a brute-force scan rather than graph search, and ships LangChain, LlamaIndex, Haystack and Agno integrations.
Who it is for: Engineers embedding compressed vector search in Python or Rust apps
Strengths
- README reports 4.4× faster than FAISS FastScan at 4-bit, at matching recall
- 7.8× smaller than float32 at 4-bit, 15.5× at 2-bit
- No training or rebuild; add and remove cost microseconds per vector
- Allowlist filtering and crash-safe incremental sync(); drop-in stores for four frameworks
Weaknesses
- Scans every vector, so query cost grows linearly with corpus size
- Lossy: top-1 matches float ground truth 96–98% at 4-bit on OpenAI embeddings
- Library only: no network API, replication or multi-tenant service
- Dimension must be a multiple of 8 and at most 16,384; only 2 or 4 bits
What it needs
- no GPU
Also in Vector databases
See all 12| Rank | Project | Score |
|---|---|---|
| 1 | MilvusDistributed vector database with dense, sparse and hybrid search at scale | 82 out of 100 |
| 2 | MeilisearchRust search engine API with full-text, vector and hybrid search | 75 out of 100 |
| 3 | QdrantRust vector database with payload filtering, REST and gRPC | 69 out of 100 |
| #4 | faissC++ library for similarity search and clustering of dense vectors | 66 out of 100 |
| #5 | ChromaEmbedding database with a four-function API for Python and JavaScript | 62 out of 100 |