[Model Request] FastEmbed new model request: dunzhang/stella_en_1.5B_v5

[richard-deetlefs]

Good day,

Would you please add dunzhang/stella_en_1.5B_v5 to FastEmbed? It is not thaaat big (5.75 GB), but has an excellent MTEB retrieval score and MIT license.

I am also a paid qdrant cloud user if that helps with the argument ;)

Thanks!

[PylotLight]

@joein Any possibility of getting this one added?

[joein]

Hey @richard-deetlefs @PylotLight
Sorry for the late response

We might not be able to add this model in the closest time, however, as of fastembed v0.6.0 we support adding custom models to fastembed in runtime

As I can see, the model providers have converted it to onnx, so should not be a problem, unless the model follows the typical preprocessing/postprocessing (just pooling / normalization) steps

An example of adding a custom model

In this particular case, you would also need to set additional_files=["onnx/model.onnx_data"]

[PylotLight]

Got it. Thanks.

[richard-deetlefs]

Great! Thanks!

Do you have plans to do the same for Sparse text embeddings, Late interaction model, and the others?

[joein]

Eventually, yes, we'd like to add this feature to the other models as well

However, it might not be reasonable for some of the models because they have too specific steps in preprocessing/postprocessing which are impossible to reuse

Image models are the most probable candidates to get the custom models support feature next