Repository navigation
Qdrant/bm25 refetches from the Hub when fully cached (offline probe ignores the mock model_file) #675
Description
What happened: On fastembed >=0.7.4, loading Qdrant/bm25 with a fully populated cache (HF_HUB_OFFLINE=1 or local_files_only=True) still triggers a live Hugging Face Hub fetch instead of resolving offline.
Why: bm25 is a mock model with model_file "mock.file" (no real weight file). The offline cache probe in ModelManagement.download_model requires the model_file to exist on disk; "mock.file" is never downloaded, so the check is always False and bm25 falls through to the network even when every real additional_files stop-word list is cached. #577 fixed the general cached case but not this mock-model path.
Repro: cache Qdrant/bm25, set HF_HUB_OFFLINE=1, load it. A Hub request is still attempted.
Activity
joein commentedon Oct 2, 2026
Thanks for the detailed report. This is fixed on
mainby #728.Qdrant/bm25now uses the repo'sconfig.jsonas itsmodel_fileinstead ofmock.file, so the cache check finds a file that's actually on disk. #728 also removedtamil.txtfromadditional_files. The Hub repo doesn't have that file, so listing it would have kept failing the check.I tested it with a fully cached model and all network connections blocked. On
main,Qdrant/bm25loads from the cache with no network requests in all three modes:HF_HUB_OFFLINE=1,local_files_only=True, and the defaults. Before #728, the first two failed withCould not load model Qdrant/bm25 from any source., and the default mode went to the Hub even though the cache was full.The fix isn't in 0.8.1 and will ship in the next release.