"Error while generating embeddings: runtime error: loading model failed: specified file not found in archive" Huggingface /settings endpoint about meilisearch

[sanders41]

The issue may be that a BERT model is required, I don't think is BAAI/bge-m3 BERT? Have you tried BERT models and still get the same error?

https://www.meilisearch.com/docs/learn/experimental/vector_search#generate-auto-embeddings-with-huggingface

[emraza1ai]

You are right. BAAI/bge-m3 is based on XLM-Roberta not BERT.
But I have tried avsolatorio/GIST-all-MiniLM-L6-v2 which is based on BERT to no avail:

uid=73 index_uid='meeting_actions' status='failed' type='settingsUpdate' details={'embedders': {'default': {'source': 'huggingFace', 'model': 'avsolatorio/GIST-all-MiniLM-L6-v2'}}} error={'message': 'internal: Error while generating embeddings: runtime error: loading model failed: specified file not found in archive.', 'code': 'internal', 'type': 'internal', 'link': 'https://docs.meilisearch.com/errors#internal'} canceled_by=None duration='PT23.760405S' enqueued_at=datetime.datetime(2024, 2, 10, 14, 34, 31, 113366) started_at=datetime.datetime(2024, 2, 10, 14, 34, 31, 115745) finished_at=datetime.datetime(2024, 2, 10, 14, 34, 54, 876150)

[emraza1ai]

I even tried the model mentioned at doc link shared by you @sanders41 https://www.meilisearch.com/docs/learn/experimental/vector_search#generate-auto-embeddings-with-huggingface,
it downloads BAAI/bge-base-en-v1.5 :

config.json [00:00:00] [███████████████████████████████████████████████████████████████████] 777 B/777 B 1.09 KiB/s (0s)
tokenizer.json [00:00:01] [████████████████████████████████████████████████████] 694.72 KiB/694.72 KiB 459.47 KiB/s (0s)
pytorch_model.bin [00:01:36] [███████████████████████████████████████████████████] 417.71 MiB/417.71 MiB 4.32 MiB/s (0s)

But then it still gives the same error:

uid=75 index_uid='meeting_actions' status='failed' type='settingsUpdate' details={'embedders': {'default': {'source': 'huggingFace', 'model': 'BAAI/bge-base-en-v1.5'}}} error={'message': 'internal: Error while generating embeddings: runtime error: loading model failed: specified file not found in archive.', 'code': 'internal', 'type': 'internal', 'link': 'https://docs.meilisearch.com/errors#internal'} canceled_by=None duration='PT102.462092400S' enqueued_at=datetime.datetime(2024, 2, 10, 14, 53, 38, 266617) started_at=datetime.datetime(2024, 2, 10, 14, 53, 38, 304698) finished_at=datetime.datetime(2024, 2, 10, 14, 55, 20, 766790)

[sanders41]

I was able to recreate this on Linux so it isn't a Windows specific issue. Hugging Face is giving a 401 access denied error. So it looks like a token is needed. hf-hub looks to have a way to pass a token (https://docs.rs/hf-hub/0.3.2/hf_hub/api/sync/struct.ApiBuilder.html#method.with_token), but I don't see a way to pass the token to Meilisearch.

uid=9 index_uid='movies' status='failed' task_type='settingsUpdate' details={'embedders': {'default': {'source': 'huggingFace', 'model': 'bge-base-en', 'revision': 'v1.5', 'documentTemplate': "A movie titled '{{doc.title}}' whose description starts with {{doc.overview|truncatewords: 20}}"}}} error={'message': 'internal: Error while generating embeddings: error: fetching file from HG_HUB failed: request error: https://huggingface.co/bge-base-en/resolve/v1.5/config.json: status code 401.', 'code': 'internal', 'type': 'internal', 'link': 'https://docs.meilisearch.com/errors#internal'} canceled_by=None duration='PT0.098242910S' enqueued_at=datetime.datetime(2024, 2, 10, 16, 53, 50, 198972) started_at=datetime.datetime(2024, 2, 10, 16, 53, 50, 203195) finished_at=datetime.datetime(2024, 2, 10, 16, 53, 50, 301438)

[emraza1ai]

You are facing a different error:

error={'message': 'internal: Error while generating embeddings: error: fetching file from HG_HUB failed: request error: https://huggingface.co/bge-base-en/resolve/v1.5/config.json: status code 401.'

This is because I think you didn't specify BAAI/ before bge-base-en-v1.5, therefore it is resolving to https://huggingface.co/bge-base-en/resolve/v1.5/config.json

Correct way:

curl -X PATCH "http://localhost:7700/indexes/meeting_actions/settings" -H "Content-Type: application/json" --data-binary "{"embedders": {"default": {"source": "huggingFace", "model": "BAAI/bge-base-en-v1.5"}}}"

Wrong way:

curl -X PATCH "http://localhost:7700/indexes/meeting_actions/settings" -H "Content-Type: application/json" --data-binary "{"embedders": {"default": {"source": "huggingFace", "model": "bge-base-en" , "revision": "v1.5"}}}"
OR
curl -X PATCH "http://localhost:7700/indexes/meeting_actions/settings" -H "Content-Type: application/json" --data-binary "{"embedders": {"default": {"source": "huggingFace", "model": "bge-base-en-v1.5"}}}"

[sanders41]

I tried both ways and got the same error, but I'll try again to make sure I didn't have a typo.

[sanders41]

Apparently I did have a typo, I tried again and didn't have any errors this time.

[emraza1ai]

Glad for you!
But coming back to my original error:

uid=69 index_uid='meeting_actions' status='failed' type='settingsUpdate' details={'embedders': {'default': {'source': 'huggingFace', 'model': 'BAAI/bge-base-en-v1.5'}}} error={'message': 'internal: Error while generating embeddings: runtime error: loading model failed: specified file not found in archive.', 'code': 'internal', 'type': 'internal', 'link': 'https://docs.meilisearch.com/errors#internal'} canceled_by=None 

What should I do? I looked around that maybe it might be symlink issue in windows. But no I don't think so, seems to be some other problem.

[emraza1ai]

Any update on this?

[dureuill]

Hello 👋, and thank you for your report ❤️

Given that the issue doesn't reproduce on Linux or macOS, it might be Windows-specific.

👉 Windows bugs are not the priority for the Meili team, so we need someone from the community to check if they can reproduce the bug on their machine, and then, someone to fix the bug.

That said, I'll try to reproduce it on a Windows machine if I can find the time.