search

qdrant / fastembed

Fast, Accurate, Lightweight Python library to make State of the Art Embedding
https://qdrant.github.io/fastembed/
Apache License 2.0
1.37k stars 99 forks source link

[Bug/Model Request]: Load model files from path, not from huggingface cach directory #321

Open satyaloka93 opened 1 month ago

satyaloka93 commented 1 month ago

What happened?

Unable to test this in my organization, as we do not use hugging face cache folders for models, models are downloaded via git, scanned, then allowed for usage. I see some attempt to use local files via 'local_files_only' kwarg in this PR, but this won't work apparently as I do not have files in the snapshot format. Request loading models from a normal directory, like transformers/sentence-transformers and most other frameworks. Really would like to incorporate this technology in our information retrieval, but this is a show stopper.

What Python version are you on? e.g. python --version

Python 3.10

Version

0.2.7 (Latest)

What os are you seeing the problem on?

Linux

Relevant stack traces and/or logs

No response

joein commented 1 month ago

Hi @satyaloka93

You want to put the same files as in HF hub to the cache directory and initialize from them. We are not talking about some custom models / files right now, right?

satyaloka93 commented 1 month ago

Hi, they are the files from the Qdrant HF repo: https://huggingface.co/Qdrant/Splade_PP_en_v1/tree/main. Our organization pulls them via git, scans, and moves them where we can load them up. When I try to load from that directory, even using it as cache_dir and local_files_only=True, it does not work. I’m assuming because it’s expecting to have a cache structure, versus the normal HF files in your repo.