Clean-room Zig reimplementation reproduces published IVFPQ memory-compression numbers

[SMC17]

Hi FAISS maintainers + community,

Posting a complementary-reproduction note. I've been building a Zig 0.16 reimplementation of the four FAISS index families that matter most for general use — Flat, HNSW, IVFFlat, IVFPQ — without consulting your C++ source (paper-only implementation).

Repo: SMC17/faiss-zig (AGPL-3.0)

At the default canonical config (nlist=100, nprobe=10, M=8, D=128), our IVFPQ measures:

• 16.94x total memory compression vs raw f32 storage (295 KB total vs 5,000 KB at N=10,000)
• 42.67x per-vector compression (12 bytes/vec vs 512)
• Recall @ 10: 0.10 on uniform [0,1]^128 random data (expected to recover toward your published 0.6–0.8 on real clustered embeddings; we have not yet bench'd on real embeddings)

These numbers track your published documentation — posting in case the reproduction signal is useful, and because community contributors implementing FAISS-equivalent libraries in non-C++ languages benefit from the data point.

Throughput at default config: 522 μs/q, 1,915 q/s, 3.43x vs our own FlatIndex baseline at this N.

Honest caveats:

• Single-platform tested (x86_64-linux, Zig 0.16 ReleaseFast)
• No GPU
• No OPQ rotation, no k-means++ init (both v0.8 roadmap)
• Cosine / inner-product only for FlatIndex at v0.7; HNSW / IVFFlat / IVFPQ are L2-squared only

Not a competing implementation — explicitly a Zig-side substrate for projects that need an in-process vector store without the Python boundary. Our README cites FAISS as the canonical reference.

Thanks for the canonical work.