Datasets:

Final-Progs
/

Semantic-Search-Engine-with-Vectorized-DB

license: mit
task_categories:
  - feature-extraction
tags:
  - vector-database
  - semantic-search
  - ivf-index
  - embeddings
  - information-retrieval
size_categories:
  - 10M<n<100M

Semantic Search Engine with Vectorized DB — Artifacts

This repository hosts the pre-computed on-disk index artifacts for the 20,000,000 vector database (OpenSubtitles_en_20M_emb_64.dat), built for the Advanced Database Systems project (Cairo University, Faculty of Engineering).

📁 Repository Structure

semantic-search-artifacts/
│
├── README.md                          # Repository documentation & usage guide
│
├── production/
│   ├── m1_ivf_k4096/                  # Module 1: IVF coarse index & cluster-sorted vectors
│   │   ├── centroids.f32              # (4096, 64) float32 L2-normalized coarse centroids
│   │   ├── offsets.i64                # (4097,) int64 inverted list boundary offsets
│   │   ├── ids.i32                    # (20,000,000,) int32 original row IDs (cluster-sorted)
│   │   ├── vectors.f32                # (20,000,000, 64) float32 normalized vectors (cluster-sorted)
│   │   ├── labels.i16                 # (20,000,000,) int16 cluster labels for original rows
│   │   ├── meta.json                  # Complete build parameters, stage timestamps & invariants
│   │   └── SHA256SUMS.txt             # Cryptographic SHA-256 checksums
│   │
│   └── m2_pq/                         # Module 2: Product Quantization (in progress)
│       ├── pq_codes.*
│       ├── pq_codebooks.*
│       ├── pq_meta.json
│       └── SHA256SUMS.txt
│
└── reports/
    ├── full_index_report.json         # Complete validation & recall metrics for 20M index
    └── dev_index_report.json          # Validation report for 1M dev index

📊 Module 1 Production Index Details (`m1_ivf_k4096`)

Total Vectors ($N$): 20,000,000
Vector Dimension ($D$): 64 (float32)
Clusters ($K$): 4096
Clustering Method: MiniBatchKMeans on 2M L2-normalized vector sample, final centroids L2-normalized.
Empty Clusters: 0 (min size = 680, max size = 22,911, mean = 4,882.8)
Coverage@64 (Upper Bound Recall): 99.30% (exceeds team target of $\ge 99.0%$)
Build Time: 304 seconds (~5 minutes)

🔒 Verification & Checksums

File	Shape / Dtype	Size	SHA-256 Checksum
`centroids.f32`	`(4096, 64)` `float32`	1,048,576 B (1.0 MB)	`deb0e2e88c764503ddbcf5eaf68d3807497228d300560477abd3fe642c635fff`
`offsets.i64`	`(4097,)` `int64`	32,776 B (32 KB)	`7894cf57c10e8a4fa25aab62ccb6c4c4a8a08d4bc744091c5904fdefb1d237b4`
`ids.i32`	`(20000000,)` `int32`	80,000,000 B (80 MB)	`6675a2ac22afbec3f65b467d5f3663d4618e541293ddba4ee10d3940c6035cab`
`labels.i16`	`(20000000,)` `int16`	40,000,000 B (40 MB)	`d3294a967e330ad2ce757ff15abe4938274315b9fd6a9b43c08d9480f7c1d1b5`
`meta.json`	JSON metadata	3,407 B (~3.4 KB)	`b86fe06ba30cf377994f271908958de6fd618ea6b818caa24102ecf648515991`
`vectors.f32`	`(20000000, 64)` `float32`	5,120,000,000 B (5.12 GB)	`caaadbd8454542fbd7c329314134730ef79fd7b228130027b60f20b2a7476054`

💻 Download & Integration Instructions

In Python using `huggingface_hub`

from huggingface_hub import hf_hub_download

# Example: Download centroids and offsets for Module 3 (Retrieval)
centroids_file = hf_hub_download(
    repo_id="Abdelrahman610/Semantic-Search-Engine-with-Vectorized-DB",
    filename="production/m1_ivf_k4096/centroids.f32",
    repo_type="dataset",
)

offsets_file = hf_hub_download(
    repo_id="Abdelrahman610/Semantic-Search-Engine-with-Vectorized-DB",
    filename="production/m1_ivf_k4096/offsets.i64",
    repo_type="dataset",
)

Download the entire index folder via CLI

huggingface-cli download Abdelrahman610/Semantic-Search-Engine-with-Vectorized-DB     --repo-type dataset     --local-dir ./downloaded_artifacts