jina-embeddings-v3 model

[yrayegan]

Model description

Support for jina-embeddings-v3

Prerequisites

• The model is supported in Transformers (i.e., listed here)
• The model can be exported to ONNX with Optimum (i.e., listed here)

Additional information

No response

Your contribution

None

[nemphys]

+1

[nemphys]

@xenova I managed to use the ONNX model from huggingface with a few minor tweaks, but I don't know how you would like to handle them:

1. The model config.json needs the "model_type": "xlm-roberta" property, otherwise its type is not properly detected by transformers.js
2. The model_inputs during model execution need a 3rd parameter which is the task_id index from the lora_adaptations list ("retrieval.query", "retrieval.passage", "separation", "classification", "text-matching")
3. The output results property is "text_embeds", which needs to be added to the relevant inspection list.

1/3 are easy, but for 2 I suppose we need to create special classes/parameters.

Also, there is the Matryoska embeddings thing; does this imply mere dimension clamping of the results or is it something more than that that needs special handling?