Cannot specify model version for SentenceTransformer embedding retriever

[laffeychris]

Describe the bug
Sentence transformer models do not allow version to be specified.

Error message
When passing a model_version argument to nodes.EmbeddingRetriever, it ignores this and only downloads the latest version of the sentence-transformer model from huggingface (or from cache).

Expected behavior
Download the specified version.

Additional context
This is due to not propagating the revision argument to SentenceTransformer(), via _SentenceTransformersEmbeddingEncoder

Here, we should have passed the version:

self.embedding_model = SentenceTransformer(
    retriever.embedding_model, 
    device=str(retriever.devices[0]), 
    use_auth_token=retriever.use_auth_token, 
    revision=retriever.model_version,
)

To Reproduce

from haystack import nodes


nodes.EmbeddingRetriever(
    embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
    model_version="39c51f707d95c3c40f68b47f1df9ae52319da356",
    model_format="sentence_transformers",
)

Check cache to see what models were downloaded:

ls -al ~/.cache/huggingface/hub/models--sentence-transformers--multi-qa-mpnet-base-dot-v1/snapshots
>> 3af7c6da5b3e1bea796ef6c97fe237538cbe6e7f  # latest version, not specified version

FAQ Check

• Have you had a look at our new FAQ page?

System:

• OS: Mac
• GPU/CPU: both
• Haystack version (commit or version number): 1.26
• DocumentStore: N/A
• Reader: N/A
• Retriever: EmbeddingRetriever

[laffeychris]

I am happy to open an MR for this

[laffeychris]

@masci - could we release this change in a fix (v1.26.3)?
I have an internal issue that this would solve. 🙏

[masci]

@laffeychris the release candidate is out if you want to give it a spin https://pypi.org/project/farm-haystack/1.26.3rc1/ - I'm planning to release the final artifact tomorrow after some testing.