Hi @Denaldo,
Our API Inference supports multiple tasks. For certain models, we provide a straightforward abstraction for embedding similarity, such as with sentences. The tag and/or pipeline_tag establishes the correct task on the API Inference backend for all compatible models on our hub.
When using sentence-similarity, the backend establishes a sentence similarity pipeline. It expects multiple sentence inputs, which will subsequently be transformed into embeddings and compared through cosine similarity
When the model is set for feature-extraction, it expects the input sentence and returns the corresponding embeddings vector.
So if you need a model that is not supported into the feature-extraction pipeline, you can duplicate it and set the correct tag and pipeline_tag.
You can check the README.md in different models for comparison