Turbovec

turbovec — local vector index

Rust vector index on Google's TurboQuant quantizer. 10M float32 docs (31GB) fit in ~4GB. Faster than FAISS. No train step: add vectors, they are indexed. Fully local — nothing leaves the machine.

Where it is installed (v0.7.0, per host)

Run scripts with that interpreter; do not pip-install elsewhere.

API (v0.7.0 — differs from upstream README)

import numpy as np
from turbovec import TurboQuantIndex

idx = TurboQuantIndex(dim=1536, bit_width=4)
idx.add(vecs)                      # 2D float32, C-contiguous: np.ascontiguousarray(x.astype(np.float32))

scores, ids = idx.search(q, k=10)  # q must be 2D (batch): vecs[7:8], not vecs[7]
                                   # returns 2D arrays (one row per query)

mask = np.zeros(len(idx), dtype=bool)   # filtered search: bool mask over slots
mask[allowed_slots] = True
scores, ids = idx.search(q, k=10, mask=mask)

idx.write("index.tq")              # persistence round-trips exactly
idx = TurboQuantIndex.load("index.tq")

Gotchas:

• Queries MUST be 2D; a 1D vector raises TypeError: 'ndarray' object cannot be cast as 'ndarray'.
• mask is a bool array of length len(idx) (slot-based), not an id allowlist. The README's IdMapIndex.add_with_ids/allowlist= API is newer than the 0.7.0 wheel.
• bit_width=4 is the standard memory/recall tradeoff; pair with any local embedding model for an air-gapped RAG stack.

When to use

• Semantic search over local corpora (memory, knowledge base, transcripts)
• RAG retrieval where privacy/latency matters (no Pinecone round-trip)
• Hybrid retrieval: narrow candidates with SQL/BM25 first, then dense rerank via mask