You could just exclude all those fields from being passed to the LLM when indexing and then add them manually in the way you desire to the prompt sent to the LLM at query time
Just to be sure we talk about the same thing: I'm not talking about metadata keys, but the templates that transform a node to a piece of text within the context's prompt. If I follow you, best option for my use case would be a node postprocessor to override templates if necessary. Sounds good indeed
