v2.1.1 — Reproducible dependencies

Reproducible dependencies

This patch release pins the Python dependencies so that a fresh install reproduces the published retrieval behaviour.

Fixed

• Pinned the embedding stack in requirements.txt: torch 2.11.0, transformers 5.9.0, sentence-transformers 5.3.0, tokenizers 0.22.2, faiss-cpu 1.13.2, numpy 2.4.4, huggingface_hub 1.17.0, safetensors 0.7.0, accelerate 1.13.0, einops 0.8.2.
  Unpinned newer releases loaded Qwen3-Embedding in bfloat16. The resulting numerical drift moved some queries (e.g. the GCD benchmark query) into a neighbouring IVF cluster, so they returned unrelated code. With the pins, all 27 evaluation queries again match the published results.

Added

• CI post-deploy smoke test: after every deploy, the workflow checks that the GCD query returns GCD implementations, not just that /health answers.

Documentation

• Explicit file modes (664 files, 775 directories) for the model and index uploaded to Toolforge, with a warning that scp -p from Windows preserves world-writable 0666 modes.
• update_release.py now also updates the SWHID link of the README badge.

Unchanged

The code, FAISS index, SQLite database and model described in the SoftwareX article (10.1016/j.softx.2026.103127, release v2.1.0) are unchanged.

Software Heritage: swh:1:dir:cf7b3a2538479f6a302018b15c0725d84977520e

Full changelog: v2.1.0...v2.1.1