You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
feat: bundle retrieval models into an offline wheel for air-gapped installs
Ship the embedding (BAAI/bge-small-en-v1.5) and rerank
(ms-marco-MiniLM-L-12-v2) models inside the package so a pip-installed
wheel runs with no network access. cfgload transparently falls back to
the bundled models (offline mode) when the configured cache lacks them,
while still respecting a user-provided cache that already has them.
The model-bundled wheel (~260 MB) exceeds public PyPI's 100 MB/file
limit, so it is built in a dedicated release job and attached to the
GitHub Release for download into a self-hosted index (Artifactory); the
lean wheel continues to publish to PyPI unchanged. The offline wheel is
built against chunksilo-2.3.3-linux-x86_64.lock (pip --no-deps) so the
dependency closure is identical to v2.3.3 with no upgrades.
- cfgload: _resolve_bundled_models() picks the packaged cache + offline
- pyproject: package-data includes _bundled_models, excludes HF blobs/
- manual-release: build_offline_wheel job downloads/dereferences models
- add v2.3.3 fully-pinned transitive lock; gitignore the bundle dir
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
0 commit comments