Tamaneena β€” Quran semantic search assets

Four files that let the Tamaneena app answer a question about the Quran by meaning, entirely on the device and with no network after the first download. They are fetched once, verified against digests pinned in the app's source, and never sent anywhere.

Files

Release directory eg2-text-270m-q4-v3/. The names are contractual, not cosmetic: model_q4.onnx_data is the name recorded inside the graph, and the app writes each file under its own name beside the others.

file bytes SHA-256
quran_semantic_index.qsi 4,840,504 e1cf4ca41ec75c9dc8133dda45adab4994d0ebe40d3fce43bab598434c15cb3b
quran_semantic_tokenizer.qstok 8,393,416 6a5007261c905bf443fb705b1ba1f7990f1ec3c4edf540e20b5f29126d06c0c9
quran_semantic_model.onnx 480,736 99e95af720ed06d939204c1b446f44f51fa3f0eb0598a3aaa5043e076e22d75a
model_q4.onnx_data 174,028,800 c3975f2d1ab7a1878ae31a7d7a9b7804a827aff3800b60dfceafce21cac3df49

Total 187,743,456 bytes (179 MB). A file whose digest does not match is dropped by the app rather than used.

Model

EmbeddingGemma-2 text tower, ONNX q4, Apache-2.0.

Source onnx-community/embeddinggemma-2-ONNX, revision daa72c51243991dfcaf9f9137d2c573d8f7790c0 β€” pinned, not main.

quran_semantic_model.onnx is that revision's onnx/model_q4.onnx with the 33 nodes of /model/multimodal_merge/ removed and their consumers rewired to the token embeddings, which drops the image_features, video_features and audio_features inputs. Outputs are bitwise identical on 137 measured texts. The weights are untouched: model_q4.onnx_data is redistributed byte for byte. quran_semantic_tokenizer.qstok is that revision's tokenizer.json in a packed binary layout; token ids are identical on 516 measured texts.

The index holds no Quran text

quran_semantic_index.qsi contains not a verse and not a word β€” measured, there is not one run of two consecutive Arabic characters anywhere in it. It is 6,236 int8 vectors of 768 dimensions with one scale per vector and one canonical verse number per row. It cannot be read without verse text supplied from outside it.

The vectors were computed from Tanzil Quran Text (Simple Clean, Version 1.1), licensed Creative Commons Attribution 3.0, from the Tanzil Project β€” https://tanzil.net. The copyright block below is reproduced verbatim, and is also embedded inside the .qsi file itself so the derived file carries it even when fetched alone.

# PLEASE DO NOT REMOVE OR CHANGE THIS COPYRIGHT BLOCK
#====================================================================
#
#  Tanzil Quran Text (Simple Clean, Version 1.1)
#  Copyright (C) 2007-2026 Tanzil Project
#  License: Creative Commons Attribution 3.0
#
#  This copy of the Quran text is carefully produced, highly
#  verified and continuously monitored by a group of specialists
#  at Tanzil Project.
#
#  TERMS OF USE:
#
#  - Permission is granted to copy and distribute verbatim copies
#    of this text, but CHANGING IT IS NOT ALLOWED.
#
#  - This Quran text can be used in any website or application,
#    provided that its source (Tanzil Project) is clearly indicated,
#    and a link is made to tanzil.net to enable users to keep
#    track of changes.
#
#  - This copyright notice shall be included in all verbatim copies
#    of the text, and shall be reproduced appropriately in all files
#    derived from or containing substantial portion of this text.
#
#  Please check updates at: http://tanzil.net/updates/
#
#===================================================================

Retrieval, measured

Index form imlaei, field ayah only, no tafsir rows. On a 64-query topical benchmark and a 12-query literal one:

ranking H@1 H@10
lexical only (topical) 0.0 0.0
semantic + hybrid 20.3 50.0
+ stem-overlap reranker 29.7 78.1
literal queries, throughout 100.0 100.0

The reranker needs nothing on disk: it reads the verse text the app already ships.

License summary

  • Engine and weights: Apache-2.0 (EmbeddingGemma-2).
  • Verse text the vectors were computed from: CC BY 3.0 (Tanzil Project), attributed above and in NOTICE.
  • No access gate and no private repository: the app downloads without a token.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for tamaneena/quran-semantic

Quantized
(2)
this model