Maktabati/shamela-bm25

Shamela BM25 Index (SQLite FTS5)

BM25 full-text search index for Maktabati/shamela-vectors, covering the complete Al-Maktaba Al-Shamela corpus — 8,589 classical Islamic books.

Designed as the lexical retrieval component of a hybrid search pipeline: BM25 (this repo) + Dense Embeddings (shamela-vectors) → RRF Fusion

---

Contents

File	Description
bm25_shamela.db	SQLite FTS5 database (~3GB)
build_bm25_index.py	Script to rebuild the index from scratch
requirements.txt	Python dependencies

hybrid_search.py (RRF fusion pipeline) is maintained separately and links both this index and shamela-vectors.

---

Why BM25 + Dense?

Dense embeddings (semantic search) excel at conceptual queries but miss exact terminology. Classical Islamic texts present a specific challenge: users searching in Modern Standard Arabic often miss Classical Arabic variants.

Query	Dense only	BM25 only	Hybrid
فائدة بنكية (bank interest)	finds semantic matches	misses ربا if not exact	finds both
حكم الغرر (ruling on uncertainty)	good	good	best
سند specific narrator name	poor	excellent	excellent

---

Arabic Preprocessing

The index applies the following normalization before indexing (consistent with shamela-vectors):

# Diacritics removed
# أإآٱ → ا
# ة    → ه
# ى    → ي

This ensures robust recall across orthographic variants common in classical Arabic manuscripts.

---

Corpus Stats

Metric	Value
Books	8,589
Categories	40
Chunks indexed	~4M
Chunk size	512 tokens
Overlap	50 tokens
Source	AuthenticIlm/Shamela4_Full_DB

---

Schema

Each FTS5 row contains:

text_norm     → normalized Arabic text (indexed, searchable)
point_id      → UUID matching Maktabati/shamela-vectors exactly
book_id       → Shamela book ID
title         → Arabic book title
author        → Arabic author name
page          → Page reference (OpenITI-compatible format, e.g. V02P045)
category_ar   → Arabic category name

The point_id field is a deterministic UUID computed identically to shamela-vectors, enabling direct cross-index lookup without a join table.

---

Usage

Option A — Download pre-built index

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Maktabati/shamela-bm25",
    filename="bm25_shamela.db",
    repo_type="dataset"
)

Option B — Build from scratch

# Requires: AuthenticIlm/Shamela4_Full_DB downloaded locally
pip install -r requirements.txt
python build_bm25_index.py --root /path/to/shamela_hf

Basic BM25 search

import sqlite3

def bm25_search(query: str, db_path: str, limit: int = 20):
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    rows = conn.execute("""
        SELECT point_id, title, author, page, book_id
        FROM   chunks
        WHERE  text_norm MATCH ?
        ORDER  BY rank
        LIMIT  ?
    """, (query, limit)).fetchall()
    conn.close()
    return [dict(r) for r in rows]

results = bm25_search("حكم الصلاة بغير وضوء", "bm25_shamela.db")

---

Hybrid Search

For full hybrid retrieval (BM25 + Dense + RRF), use this index together with Maktabati/shamela-vectors:

Query
  ├── BM25   → SQLite FTS5 (this repo)    → Top-150
  ├── Dense  → Qdrant / shamela-vectors   → Top-150
  └── RRF Fusion                          → Top-10

RRF score: 1/(k + rank_bm25) + 1/(k + rank_dense) with k=60

---

Related

• Maktabati/shamela-vectors — Dense embeddings (Qdrant, multilingual-e5-base)
• AuthenticIlm/Shamela4_Full_DB — Source corpus

---

License

Apache 2.0. See source corpus license for text content rights.