I'm using llama index, my current embedding model is a BG, my LLM is Claude and my vector store is opensearch. I'd like to use XML tags to format the metadata before feeding it to Claude, this should be improving the LLM performance.

I can do so using Document.metadata_seperator, Document.metadata_template and Document.text_template when creating and filling up my indices, but this will also change the text format for the embedding model, which seems to decrease performance.

My question is, how can I have one specific text format for my embedding model when indexing my documents but a different one when activating my index, retrieving, and querying the LLM ?

Thank you very much