Incompatibility Between Haystack 2.4.0 and Sentence-Transformers 2.x Due to `model_kwargs` Argument

[movchan74]

Describe the bug Haystack 2.4.0 specifies a dependency on sentence-transformers version >=2.3.0. However, the SentenceTransformersTextEmbedder component uses the model_kwargs argument, which is not supported in sentence-transformers versions below 3.0. As a result, initializing the SentenceTransformersTextEmbedder with sentence-transformers 2.x (e.g., 2.7.0) leads to a TypeError.

Error message

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/root/haystack_bug/venv/lib/python3.10/site-packages/haystack/components/embedders/sentence_transformers_text_embedder.py", line 167, in warm_up
    self.embedding_backend = _SentenceTransformersEmbeddingBackendFactory.get_embedding_backend(
  File "/root/haystack_bug/venv/lib/python3.10/site-packages/haystack/components/embedders/backends/sentence_transformers_backend.py", line 37, in get_embedding_backend
    embedding_backend = _SentenceTransformersEmbeddingBackend(
  File "/root/haystack_bug/venv/lib/python3.10/site-packages/haystack/components/embedders/backends/sentence_transformers_backend.py", line 66, in __init__
    self.model = SentenceTransformer(
TypeError: SentenceTransformer.__init__() got an unexpected keyword argument 'model_kwargs'

Expected behavior The SentenceTransformersTextEmbedder should initialize correctly without raising an exception. The dependency on sentence-transformers should either be fixed to ensure compatibility with the model_kwargs argument (i.e., sentence-transformers>=3.0.0) or the code should be adjusted to work with sentence-transformers 2.x.

Additional context This issue occurs because the current dependency in the pyproject.toml file allows the installation of sentence-transformers versions that do not support the model_kwargs argument. This bug can be resolved by either upgrading the dependency to sentence-transformers>=3.0.0 or removing the use of model_kwargs in Haystack 2.4.0.

To Reproduce

1. Install sentence-transformers version 2.x (e.g., 2.7.0) and haystack-ai version 2.4.0:

   pip install sentence-transformers==2.7.0
   pip install haystack-ai==2.4.0

2. Run the following code:

   from haystack.components.embedders import SentenceTransformersTextEmbedder
   text_embedder = SentenceTransformersTextEmbedder()
   text_embedder.warm_up()

3. Observe the TypeError due to the unsupported model_kwargs argument.

FAQ Check

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

System:

• OS: Ubuntu 22.04.3 LTS
• Haystack version: 2.4.0
• Python version: 3.10
• sentence-transformers version: 2.7.0

Suggested Fix Update the pyproject.toml to require sentence-transformers>=3.0.0 to ensure compatibility with the model_kwargs argument. Alternatively, modify the code in Haystack 2.4.0 to avoid using model_kwargs when working with sentence-transformers versions below 3.0.

[julian-risch]

Thank you @movchan74 for bringing this to our attention! We will discuss it in our team in the next sprint planning.

[anakin87]

I would say that sentence-transformers is not a core dependency of haystack-ai (we wrap imports in a lazy import block). Instead, it is a test dependency (part of the [tool.hatch.envs.test] block in pyproject.toml).

For this reason, we can't force users to install a specific version. However, we should

• update the test dependency pin
• update the LazyImport message to reflect that

[anakin87]

done in #8295, https://github.com/deepset-ai/haystack-tutorials/pull/345, and https://github.com/deepset-ai/haystack-cookbook/pull/113