Maktabati/openiti-vectors

OpenITI Vector Database — maktabati.ai

---

🇬🇧 English

This dataset contains the fully vectorized OpenITI RELEASE 2025-1-9 collection of classical Islamic texts, prepared for semantic search (RAG).

Each entry represents a text chunk from one of 8,943 works in the OpenITI corpus, together with its embedding vector and complete metadata.

Statistics:

• 4,696,703 chunks
• 8,943 works (primary editions only, status=pri from OpenITI TSV)
• approx. 470 Parquet files (approx. 45MB each, zstd-compressed)
• Period: 1st century AH to 14th century AH
• Languages: Arabic (majority), Persian, Turkish, Urdu

🔍 Live Demo

Try the keyword search demo directly in your browser — no setup required:

👉 maktabati.ai · OpenITI Search

Keyword search only (author, title, full text). For semantic / vector search, import the dataset into Qdrant using the instructions below.

---

🇸🇦 العربية

أكبر قاعدة بيانات متجهية مفتوحة المصدر لمجموعة OpenITI من النصوص الإسلامية الكلاسيكية، مُعدَّة للبحث الدلالي والذكاء الاصطناعي التوليدي (RAG).

تضمّ هذه المجموعة 4,696,703 مقطعًا نصيًا مستخرجًا من 8,943 مصنَّفًا تمتد من القرن الأول حتى الرابع عشر الهجري، وتشمل النصوص العربية والفارسية والتركية والأردية. اقتُصر على الطبعات الأولية (status=pri) تجنبًا لتكرار المحتوى، وعند تعدد الطبعات اختيرت الأطول نصًا.

تم التضمين والفهرسة بالكامل على حاسوب شخصي دون أي خدمات سحابية، باستخدام نموذج intfloat/multilingual-e5-base بأبعاد 768 ومسافة كوساين — مما يجعل هذه المجموعة قابلة للتكرار والنشر بأدوات مفتوحة المصدر.

الإحصاءات:

• 4,696,703 مقطع نصي
• 8,943 مصنَّف (الطبعات الأولية فقط من أصل 14,000+)
• نصوص من القرن الأول الهجري حتى القرن الرابع عشر الهجري
• اللغات: العربية (الأغلب)، الفارسية، التركية، الأردية
• منجز على حاسوب شخصي — دون خدمات سحابية

---

🇩🇪 Deutsch

Die größte quelloffene Vektordatenbank des OpenITI-Korpus klassischer islamischer Texte — optimiert für semantische Suche und RAG-Anwendungen.

4.696.703 Chunks aus 8.943 Werken vom 1. bis 14. Jahrhundert AH, in den Sprachen Arabisch, Persisch, Türkisch und Urdu. Es wurden ausschließlich Primäreditionen (status=pri) indexiert, um doppelte Inhalte zu vermeiden — das reduziert den Korpus von 14.000+ auf 8.943 einzigartige Werke.

Die gesamte Einbettung und Indexierung erfolgte auf einem Consumer-PC ohne Cloud-Dienste, mit dem Modell intfloat/multilingual-e5-base (768 Dimensionen, Kosinus-Distanz) — vollständig reproduzierbar mit Open-Source-Tools.

Statistiken:

• 4.696.703 Chunks
• 8.943 Werke (nur Primäreditionen, von 14.000+)
• Zeitraum: 1. bis 14. Jahrhundert AH
• Sprachen: Arabisch, Persisch, Türkisch, Urdu
• Erstellt auf Consumer-Hardware — ohne Cloud-Dienste

🇹🇷 Türkçe

OpenITI klasik İslam metinleri koleksiyonunun en kapsamlı açık kaynaklı vektör veritabanı — anlamsal arama ve RAG uygulamaları için hazırlanmıştır.

1. ile 14. yüzyıl Hicri arasında yazılmış 8.943 eserden 4.696.703 metin parçası; Arapça, Farsça, Türkçe ve Urduca dillerinde. İçerik tekrarını önlemek için yalnızca birincil baskılar (status=pri) dizine alındı — bu, 14.000+ dosyayı 8.943 benzersiz esere indirgedi.

Tüm gömme ve dizinleme işlemi, intfloat/multilingual-e5-base modeli kullanılarak (768 boyut, kosinüs mesafesi) tüketici sınıfı bir PC'de gerçekleştirildi — bulut hizmeti kullanılmadı; açık kaynak araçlarla tamamen yeniden üretilebilir.

İstatistikler:

• 4.696.703 metin parçası
• 8.943 eser (yalnızca birincil baskılar, 14.000+ içinden)
• Dönem: 1. ile 14. yüzyıl Hicri
• Diller: Arapça, Farsça, Türkçe, Urduca
• Tüketici donanımında oluşturuldu — bulut hizmeti kullanılmadı

🇮🇷 فارسی

بزرگ‌ترین پایگاه داده برداری متن‌باز از مجموعه OpenITI متون کلاسیک اسلامی — آماده برای جستجوی معنایی و کاربردهای RAG.

۴٬۶۹۶٬۷۰۳ قطعه متنی از ۸٬۹۴۳ اثر از قرن اول تا چهاردهم هجری، به زبان‌های عربی، فارسی، ترکی و اردو. تنها نسخه‌های اصلی (status=pri) نمایه‌سازی شدند تا از تکرار محتوا جلوگیری شود — این تعداد را از ۱۴٬۰۰۰+ فایل به ۸٬۹۴۳ اثر منحصربه‌فرد کاهش داد.

تمام فرآیند جاسازی و نمایه‌سازی با مدل intfloat/multilingual-e5-base (۷۶۸ بُعد، فاصله کسینوس) روی یک رایانه شخصی انجام شد — بدون هیچ سرویس ابری؛ کاملاً قابل بازتولید با ابزارهای متن‌باز.

آمار:

• ۴٬۶۹۶٬۷۰۳ قطعه متنی
• ۸٬۹۴۳ اثر (فقط نسخه‌های اصلی، از ۱۴٬۰۰۰+)
• دوره: قرن اول تا چهاردهم هجری
• زبان‌ها: عربی، فارسی، ترکی، اردو
• ساخته‌شده روی سخت‌افزار شخصی — بدون سرویس ابری

---

Technical Details

Parameter	Value
Embedding Model	intfloat/multilingual-e5-base
Vector Dimension	768
Distance Metric	Cosine
Chunk Size	512 tokens
Overlap	50 tokens
Tokenizer	XLM-RoBERTa (via multilingual-e5-base)
Passage Prefix	passage: (indexing)
Query Prefix	query: (search)

• Indexed and embedded on a consumer-grade home PC:
  • CPU: AMD Ryzen 9 5900X
  • RAM: 32GB DDR4
  • GPU: AMD Radeon RX 6900 XT (16GB VRAM)
  • Motherboard: Gigabyte X470 Aorus Ultra Gaming
  • No cloud services used; fully reproducible on consumer hardware

---

Dataset Fields

⚠ Breaking change (2026-05): Field names were standardized from a legacy mixed-language schema (German/Arabic/English) to consistent English. See Migration below.

Field	Type	Description
id	string	Deterministic SHA256-based ID
vector	float32[768]	Embedding vector
text	string	Original chunk text
text_norm	string	Normalized text (no tashkeel, unified alef variants)
sha256	string	SHA256 hash of chunk text
author	string	Arabic author name (e.g. الطبري)
title	string	Arabic book title (e.g. تاريخ الطبري)
death_year	int32	Author death year in AH (e.g. 310)
language	string	Language code (ar, fa, tr, ur)
page	string	Last OpenITI page marker before chunk (e.g. PageV01P023)
char_start	int64	Character position in source text (start)
char_end	int64	Character position in source text (end)
file_name	string	OpenITI filename (e.g. 0310Tabari.TarikhIslam.Shamela-ara1)
chunk_no	int32	Chunk number within file
indexed_at	string	ISO timestamp of indexing

---

Source Data

• OpenITI RELEASE 2025-1-9: https://github.com/OpenITI/RELEASE
• Metadata from official OpenITI_metadata_2025-1-9.tsv (author names, titles, edition status)
• Primary editions only (status=pri from TSV)
• When multiple primary editions exist: longest by char_length

Why not all 14,000+ works?

OpenITI contains multiple editions of the same work. Only primary editions (status=pri) were indexed to avoid duplicate content in search results. When multiple primary editions exist, the longest by character count was selected. This reduces the corpus from ~14,000 files to 8,943 unique works.

Usage

Import into Qdrant

import pyarrow.parquet as pq
from pathlib import Path
from qdrant_client import QdrantClient
from qdrant_client.models import (
    VectorParams, Distance, PointStruct,
    HnswConfigDiff, OptimizersConfigDiff
)

client = QdrantClient("http://localhost:6333")

# Create collection
client.create_collection(
    collection_name="openiti",
    vectors_config=VectorParams(size=768, distance=Distance.COSINE, on_disk=True),
    hnsw_config=HnswConfigDiff(m=0),  # disable HNSW during upload
    optimizers_config=OptimizersConfigDiff(indexing_threshold=0)
)

# Import Parquet files
for path in sorted(Path("openiti-vectors").glob("*.parquet")):
    table = pq.read_table(path)
    points = [
        PointStruct(
            id=row["id"],
            vector=row["vector"],
            payload={k: row[k] for k in table.schema.names if k not in ("id", "vector")}
        )
        for row in table.to_pylist()
    ]
    client.upsert(collection_name="openiti", points=points, wait=False)

# Enable HNSW after upload
client.update_collection(
    collection_name="openiti",
    hnsw_config=HnswConfigDiff(m=16, ef_construct=100, on_disk=False),
    optimizers_config=OptimizersConfigDiff(indexing_threshold=1, max_optimization_threads=4)
)
# Wait until indexed_vectors_count == points_count

Semantic Search

from transformers import AutoTokenizer, AutoModel
import torch
from qdrant_client import QdrantClient

model_name = "intfloat/multilingual-e5-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)
model.eval()

def embed(text: str) -> list:
    # Important: use "query: " prefix for search queries
    encoded = tokenizer(f"query: {text}", return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        output = model(**encoded)
    emb = output.last_hidden_state.mean(dim=1)
    return torch.nn.functional.normalize(emb, p=2, dim=1)[0].tolist()

client = QdrantClient("http://localhost:6333")

vector = embed("ما قاله ابن كثير عن يأجوج ومأجوج")
results = client.search(
    collection_name="openiti",
    query_vector=vector,
    limit=10,
    with_payload=True
)

for r in results:
    print(f"Score: {r.score:.3f} | {r.payload['author']} | {r.payload['title']}")
    print(r.payload['text'][:200])
    print()

Filter by Author

from qdrant_client.models import Filter, FieldCondition, MatchValue

results = client.search(
    collection_name="openiti",
    query_vector=vector,
    query_filter=Filter(must=[
        FieldCondition(key="author", match=MatchValue(value="ابن كثير"))
    ]),
    limit=10
)

---

Migration

If you downloaded this dataset before May 2026, the following field names have changed as part of a schema standardization (legacy mixed-language → English):

Old (German)	New (English)
autor	author
titel	title
todesjahr	death_year
sprache	language
seite	page
datei_name	file_name
chunk_nr	chunk_no
indexiert	indexed_at

To update your local Parquet files:

import pandas as pd
from pathlib import Path

RENAME_MAP = {
    "autor": "author", "titel": "title", "todesjahr": "death_year",
    "sprache": "language", "seite": "page", "datei_name": "file_name",
    "chunk_nr": "chunk_no", "indexiert": "indexed_at",
}

for path in Path("openiti-vectors").glob("*.parquet"):
    df = pd.read_parquet(path)
    df.rename(columns=RENAME_MAP).to_parquet(path, index=False)

If you have already imported into Qdrant, the recommended approach is a full re-import using the updated Parquet files — this is the cleanest and fastest path for a 4.7M point collection.

⚠ Advanced only: Updating payloads in-place via scroll + set_payload is possible but will take several hours at this scale and is not recommended. Only consider this if a full re-import is not feasible for your setup.

---

License

The texts are from the OpenITI project and subject to their respective licenses. The embeddings and metadata in this dataset are released under CC BY 4.0.

---

Citation

@dataset{maktabati_openiti_2026,
  title     = {OpenITI Vector Database},
  author    = {maktabati.ai},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/Maktabati/openiti-vectors},
  note      = {Based on OpenITI RELEASE 2025-1-9; 4,696,703 chunks embedded and indexed in 2026}
}

---

Contact

Created as part of the maktabati.ai project — semantic search in classical Islamic texts.