Add hnsw-zig: Zig-native HNSW with quantized search

[mdvsh]

Summary

• First Zig-native HNSW implementation on ann-benchmarks
• SQ4 two-phase quantized search (4-bit graph traversal, f32 rescore)
• comptime SIMD via @Vector: width=16 on x86_64 (AVX-512), width=8 on aarch64 (NEON)
• Epoch-based visited set, BFS graph reorder, multi-threaded build
• Multi-dimension support via comptime dispatch (25, 50, 100, 128, 200, 256, 384, 512, 768, 784, 960)
• 20K QPS at 0.984 recall@10 on SIFT-128 (Apple M4, single-threaded search)

Source

https://github.com/mdvsh/hnsw-zig

Validation

• Docker image builds on linux/amd64, shared library loads via ctypes
• End-to-end tested: Docker -> Python ctypes -> Zig .so -> build index -> search -> correct results
• Multi-dim verified: dim=25, 128, 256, 784 all pass end-to-end
• 64 unit tests passing (zig build test)
• CI: hnswzig added to benchmarks.yml matrix

Datasets supported

Covers all ann-benchmarks datasets: SIFT (128), GloVe (25, 50, 100, 200), Fashion-MNIST (784), GIST (960), NYTimes (256), COCO (512)

Each dimension is a separate comptime-monomorphized code path with fully unrolled SIMD. No runtime branching on dimension.

Test plan

• Docker image builds and .so loads via ctypes
• End-to-end: fit + query returns correct neighbors
• Multi-dimension: tested dim=25, 128, 256, 784
• CI matrix entry added
• Supports euclidean and angular (cosine via normalize-at-insert)