Currently, we have three, Encoder, EncoderWithSimilarity, EncoderWithConversationEncode.

Encoder and EncoderWithSimilarity can be combined, mentioning that similarity encodes are optional.

Encode conversation is trickier. A proposed solution is to introduce a Document, which is essentially a string if asked to be a string (implement a default aggregation), however, it is possible that the model can handle it as a conversation (would also solve the cases where retrieval documents can have titles and a body.