Fiqh-Embed AR/FA

Fiqh-Embed is a bilingual embedding model for semantic search over Islamic jurisprudence in the Ja'fari (Imami) tradition. It maps Arabic and Persian questions and passages into a shared 1024-dimensional space, so that a question asked in colloquial Arabic lands next to the ruling that answers it — whether that ruling is written in the Arabic of Taḥrīr al-Wasīla or the Persian of Tawḍīḥ al-Masāʾil.

The model was built by Sadiqoon Technologies from real retrieval traffic and a curated corpus of the standard reference works of the school. It continues BAAI/bge-m3 (568M parameters, 8,192-token context) and is a drop-in replacement for it: same dimensionality, same CLS pooling, no instruction prefixes.

نموذج تضمين ثنائيّ اللغة للبحث الدلاليّ في الفقه الإماميّ. يضع السؤال العربيّ — ولو بالعامّية — والنصّ الفقهيّ الذي يجيبه، عربيًّا كان أو فارسيًّا، في فضاءٍ واحد. مبنيّ على bge-m3 ويحلّ محلّه مباشرةً في أيّ منظومة استرجاع قائمة.

Highlights

• Domain-tuned retrieval. Trained on real questions people ask a fiqh assistant — short, dialectal, often imprecise — paired with the exact passages that answered them.
• Arabic ⇄ Persian. Persian source texts are retrievable from Arabic questions without translation; the two languages share the space by design.
• Drop-in for bge-m3. Identical interface and vector size; re-index and go. No query prefixes, no special pooling.
• Runs anywhere. Available as safetensors for Transformers / sentence-transformers and as GGUF for llama.cpp, Ollama and LM Studio.

Evaluation

Held-out set of 292 real questions across six categories, judged against pooled relevance labels (TREC-style, top-10 from five retrieval systems). nDCG@10:

Both models were evaluated in the same run, with the same tokenization and 384-token passage window.

Training

• Corpus: 40,183 passages from the reference works in Arabic and Persian — Ajwibat al-Istiftāʾāt, Taḥrīr al-Wasīla, al-Risāla al-Taʿlīmiyya, Risāla fī al-Ṣalāt wa-l-Ṣawm, al-Aḥkām al-Muntakhaba, Tawḍīḥ al-Masāʾil and others.
• Pairs: ~3,300 real user questions matched to their source passages by verbatim quotation and page reference, plus ~8,800 Arabic queries written for Persian passages to teach the cross-lingual mapping.
• Objective: contrastive (InfoNCE) with in-batch and mined hard negatives, Arabic and Persian mixed in every batch. Full fine-tuning of all weights.

Usage

sentence-transformers

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("sadiqoon/fiqh-embed-ar-fa")

queries = ["هل يجب الخمس في الذهب الملبوس؟"]
passages = [
    "لا يجب الخمس في ما لا يدخل في مسمّى التكسّب حتى لو مضى عليه سنة أو أكثر.",
    "طلا و جواهراتی که زن برای زینت استفاده می‌کند، در صورتی که متناسب با شأن او باشد، خمس ندارد.",
    "يستحب للمصلّي أن يأتي بالأذان والإقامة قبل الصلاة.",
]
q = model.encode(queries, normalize_embeddings=True)
p = model.encode(passages, normalize_embeddings=True)
print(q @ p.T)

Transformers

import torch
from transformers import AutoTokenizer, AutoModel
tok = AutoTokenizer.from_pretrained("sadiqoon/fiqh-embed-ar-fa")
model = AutoModel.from_pretrained("sadiqoon/fiqh-embed-ar-fa").eval()

def embed(texts):
    t = tok(texts, padding=True, truncation=True, max_length=8192, return_tensors="pt")
    with torch.no_grad():
        h = model(**t).last_hidden_state[:, 0]          # CLS pooling
    return torch.nn.functional.normalize(h, dim=-1)

GGUF

Quantized builds for llama.cpp, Ollama and LM Studio are in sadiqoon/fiqh-embed-ar-fa-gguf.

Citation

@misc{sadiqoon2026fiqhembed,
  title  = {Fiqh-Embed AR/FA: Bilingual Embeddings for Imami Jurisprudence},
  author = {Sadiqoon Technologies},
  year   = {2026},
  url    = {https://huggingface.co/sadiqoon/fiqh-embed-ar-fa}
}

License & Contact

MIT. Built and maintained by Sadiqoon Technologies Ltd, London. Questions and collaboration: info@sadiqoon.uk