vicinity 0.3.1

Approximate Nearest Neighbor Search: HNSW, DiskANN, IVF-PQ, ScaNN, quantization
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.10"
# dependencies = ["numpy", "h5py", "requests"]
# ///
"""Download and convert ann-benchmarks datasets to vicinity's binary format.

Supports the standard ann-benchmarks.com datasets (HDF5).

Usage:
    uv run scripts/download_ann_benchmarks.py sift-128-euclidean
    uv run scripts/download_ann_benchmarks.py glove-25-angular
    uv run scripts/download_ann_benchmarks.py --list

Output: data/ann-benchmarks/<name>/{train,test,neighbors}.bin (VEC1/NBR1 format)

These files can be used directly with:
    JIN_DATASET=ann cargo run --example 03_quick_benchmark --release
"""

import argparse
import os
import struct
import sys
from pathlib import Path

import h5py
import numpy as np
import requests

# Standard ann-benchmarks datasets
# Format: (name, url_suffix, distance_metric, normalize)
DATASETS = {
    "sift-128-euclidean": {
        "url": "http://ann-benchmarks.com/sift-128-euclidean.hdf5",
        "metric": "euclidean",
        "normalize": False,  # Keep original metric; use original ground truth
        "size_mb": 501,
    },
    "glove-25-angular": {
        "url": "http://ann-benchmarks.com/glove-25-angular.hdf5",
        "metric": "angular",
        "normalize": True,  # Normalize for cosine-based HNSW
        "size_mb": 121,
    },
    "glove-100-angular": {
        "url": "http://ann-benchmarks.com/glove-100-angular.hdf5",
        "metric": "angular",
        "normalize": True,
        "size_mb": 512,
    },
    "fashion-mnist-784-euclidean": {
        "url": "http://ann-benchmarks.com/fashion-mnist-784-euclidean.hdf5",
        "metric": "euclidean",
        "normalize": False,
        "size_mb": 217,
    },
    "nytimes-256-angular": {
        "url": "http://ann-benchmarks.com/nytimes-256-angular.hdf5",
        "metric": "angular",
        "normalize": True,
        "size_mb": 312,
    },
    "mnist-784-euclidean": {
        "url": "http://ann-benchmarks.com/mnist-784-euclidean.hdf5",
        "metric": "euclidean",
        "normalize": False,
        "size_mb": 217,
    },
}


def download_file(url: str, dest: Path) -> None:
    """Download with progress indicator."""
    print(f"Downloading {url} ...")
    resp = requests.get(url, stream=True)
    resp.raise_for_status()
    total = int(resp.headers.get("content-length", 0))
    downloaded = 0
    with open(dest, "wb") as f:
        for chunk in resp.iter_content(chunk_size=1024 * 1024):
            f.write(chunk)
            downloaded += len(chunk)
            if total > 0:
                pct = downloaded / total * 100
                print(f"\r  {downloaded // (1024*1024)}MB / {total // (1024*1024)}MB ({pct:.0f}%)", end="", flush=True)
    print()


def write_vec1(path: Path, vectors: np.ndarray) -> None:
    """Write vectors in VEC1 binary format."""
    n, d = vectors.shape
    with open(path, "wb") as f:
        f.write(b"VEC1")
        f.write(struct.pack("<II", n, d))
        f.write(vectors.astype(np.float32).tobytes())
    print(f"  Wrote {path}: {n} vectors x {d} dims")


def write_nbr1(path: Path, neighbors: np.ndarray) -> None:
    """Write neighbors in NBR1 binary format."""
    n, k = neighbors.shape
    with open(path, "wb") as f:
        f.write(b"NBR1")
        f.write(struct.pack("<II", n, k))
        f.write(neighbors.astype(np.int32).tobytes())
    print(f"  Wrote {path}: {n} queries x {k} neighbors")


def normalize_vectors(vectors: np.ndarray) -> np.ndarray:
    """L2-normalize vectors (required for vicinity's HNSW cosine distance)."""
    norms = np.linalg.norm(vectors, axis=1, keepdims=True)
    norms = np.maximum(norms, 1e-10)
    return vectors / norms


def recompute_ground_truth(train: np.ndarray, test: np.ndarray, k: int) -> np.ndarray:
    """Recompute ground truth using cosine distance on normalized vectors.

    Processes queries in batches to avoid O(n_test * n_train) memory spike.
    """
    print(f"  Recomputing ground truth (cosine distance, k={k})...")
    n_test = test.shape[0]
    batch_size = max(1, min(100, 500_000_000 // (train.shape[0] * 4)))  # ~500MB per batch
    neighbors = np.empty((n_test, k), dtype=np.int32)

    for start in range(0, n_test, batch_size):
        end = min(start + batch_size, n_test)
        # Cosine similarity = dot(a, b) for normalized vectors
        sims = test[start:end] @ train.T  # (batch, n_train)
        # Top-k by similarity (highest = closest)
        # Use argpartition for O(n) partial sort instead of O(n log n) full sort
        if k < sims.shape[1]:
            top_k_idx = np.argpartition(-sims, k, axis=1)[:, :k]
            # Sort the top-k by similarity (descending)
            for i in range(end - start):
                order = np.argsort(-sims[i, top_k_idx[i]])
                top_k_idx[i] = top_k_idx[i, order]
            neighbors[start:end] = top_k_idx
        else:
            neighbors[start:end] = np.argsort(-sims, axis=1)[:, :k]

        if end < n_test:
            print(f"    {end}/{n_test} queries...", flush=True)

    return neighbors


def convert_dataset(name: str, info: dict, output_dir: Path) -> None:
    """Download HDF5, convert to VEC1/NBR1."""
    output_dir.mkdir(parents=True, exist_ok=True)

    hdf5_path = output_dir / f"{name}.hdf5"
    if not hdf5_path.exists():
        download_file(info["url"], hdf5_path)
    else:
        print(f"Using cached {hdf5_path}")

    with h5py.File(hdf5_path, "r") as f:
        train = np.array(f["train"])
        test = np.array(f["test"])
        gt_neighbors = np.array(f["neighbors"])

    print(f"  Train: {train.shape}, Test: {test.shape}, GT: {gt_neighbors.shape}")

    # Normalize for cosine-based HNSW
    if info.get("normalize", False):
        print("  Normalizing vectors (L2)...")
        train = normalize_vectors(train)
        test = normalize_vectors(test)
        # Recompute ground truth on normalized vectors using cosine distance
        k = gt_neighbors.shape[1]
        gt_neighbors = recompute_ground_truth(train, test, k)

    write_vec1(output_dir / "train.bin", train)
    write_vec1(output_dir / "test.bin", test)
    write_nbr1(output_dir / "neighbors.bin", gt_neighbors)

    print(f"\nDataset ready: {output_dir}/")
    print(f"Run benchmark:")
    print(f"  cargo run --example 04_rigorous_benchmark --release -- --data-dir {output_dir}")


def main() -> None:
    parser = argparse.ArgumentParser(description="Download ann-benchmarks datasets")
    parser.add_argument("dataset", nargs="?", help="Dataset name (e.g., sift-128-euclidean)")
    parser.add_argument("--list", action="store_true", help="List available datasets")
    parser.add_argument("--output", default="data/ann-benchmarks", help="Output directory")
    args = parser.parse_args()

    if args.list or not args.dataset:
        print("Available datasets:\n")
        for name, info in DATASETS.items():
            print(f"  {name:<35} {info['metric']:<12} ~{info['size_mb']}MB")
        print(f"\nUsage: uv run {sys.argv[0]} <dataset-name>")
        return

    if args.dataset not in DATASETS:
        print(f"Unknown dataset: {args.dataset}")
        print(f"Available: {', '.join(DATASETS.keys())}")
        sys.exit(1)

    output_dir = Path(args.output) / args.dataset
    convert_dataset(args.dataset, DATASETS[args.dataset], output_dir)


if __name__ == "__main__":
    main()