Maktabati/openiti-bm25

OpenITI BM25 Index

BM25 full-text search index for the OpenITI corpus (release 2025.1.9), built on SQLite FTS5. Part of the maktabati.ai Islamic RAG pipeline.

Designed to pair with Maktabati/openiti-vectors for hybrid retrieval (BM25 + Dense + RRF fusion).

---

Statistics

Source	OpenITI 2025.1.9
Files indexed	~8,943 (one edition per work)
Languages	Arabic (ar), Persian (fa), Turkish (tr), Urdu (ur)
Chunk size	512 tokens, 50-token overlap
Tokenizer	intfloat/multilingual-e5-base

---

Files

File	Description
bm25_openiti.db.zst.aa …	SQLite FTS5 database, zstd-compressed, split into 500 MB parts
build_bm25_openiti.py	Build script — multiprocessing, resume-capable, CPU-only
requirements.txt	Python dependencies

Download size: 5,4 GB compressed · Uncompressed: 23,3 GB

Download & Reconstruct

Command line:

pip install huggingface_hub
huggingface-cli download Maktabati/openiti-bm25 --repo-type dataset --local-dir .
cat bm25_openiti.db.zst.* | zstd -d -o bm25_openiti.db

Python:

from huggingface_hub import snapshot_download
import subprocess, pathlib

local = snapshot_download("Maktabati/openiti-bm25", repo_type="dataset")
parts = sorted(pathlib.Path(local).glob("bm25_openiti.db.zst.*"))
with open("bm25_openiti.db", "wb") as out:
    subprocess.run(["zstd", "-d", "--stdout"] + [str(p) for p in parts], stdout=out)

Install zstd: sudo apt install zstd / brew install zstd

---

Schema

CREATE VIRTUAL TABLE chunks USING fts5(
    text_norm,               -- normalized Arabic text (FTS5-indexed)
    point_id   UNINDEXED,    -- UUID → matches Qdrant point IDs in openiti-vectors
    author     UNINDEXED,    -- author name (Arabic preferred, Latin fallback)
    title      UNINDEXED,    -- work title
    file_name  UNINDEXED,    -- OpenITI filename (e.g. 0310Tabari.TarikhIslam.JK000001-ara1)
    page       UNINDEXED,    -- page marker (e.g. PageV01P042)
    death_year UNINDEXED,    -- author death year (Hijri)
    language   UNINDEXED,    -- ar / fa / tr / ur
    tokenize = 'unicode61'
);

The point_id is a deterministic 32-char hex string:

sha256(f"{full_file_path}#{chunk_nr}").hexdigest()[:32]

This matches exactly the Qdrant point IDs in openiti-vectors.

---

Arabic Normalization

Applied to text_norm at index time and to queries at search time:

• Remove all diacritics (harakat, tanwin, shadda, Quranic signs …)
• أإآٱ → ا · ة → ه · ى → ي
• Remove OpenITI markup: PageV##P###, ~~, # headings
• Collapse whitespace

---

Usage

Standalone BM25

import sqlite3, re

def normalize_arabic(text):
    text = re.sub(r'[\u0610-\u061A\u064B-\u065F\u0670\u06D6-\u06DC\u06DF-\u06E4\u06E7\u06E8\u06EA-\u06ED]', '', text)
    text = re.sub(r'[أإآٱ]', 'ا', text)
    text = text.replace('ة','ه').replace('ى','ي')
    text = re.sub(r'PageV\d+P\d+|~~|#\s*', '', text)
    return re.sub(r'\s+', ' ', text).strip()

conn  = sqlite3.connect("bm25_openiti.db")
query = normalize_arabic("ما حكم الصلاة بغير وضوء")
tokens    = [w for w in query.split() if w]
fts_query = " OR ".join('"' + w.replace('"','""') + '"' for w in tokens)

rows = conn.execute(
    "SELECT point_id, author, title, file_name, page "
    "FROM chunks WHERE text_norm MATCH ? ORDER BY rank LIMIT 20",
    (fts_query,)
).fetchall()

Hybrid Search (BM25 + Dense + RRF)

RRF formula: score(d) = Σ 1 / (k + rank_i), k = 60.

---

Edition Selection

One edition per work is selected from the TSV metadata:

1. Primary edition (status = pri) preferred
2. Fallback: longest edition by character count

This matches exactly the selection in index_openiti_v3.py (Qdrant indexer), ensuring UUID consistency.

---

Rebuild

# Build index (resume-capable, multiprocessing)
python build_bm25_openiti.py

# Compress and split for upload
bash compress_split_openiti.sh

---

Relation to openiti-vectors

	openiti-bm25	openiti-vectors
Search type	Lexical (BM25)	Semantic (Dense)
Backend	SQLite FTS5	Qdrant
Model	—	intfloat/multilingual-e5-base
UUID scheme	sha256(path#chunk_nr)[:32]	identical

---

License

Text content from OpenITI — CC BY 4.0. Index structure and scripts: CC BY-SA 4.0.