If you use docling to prepare documents for input to gpt models using RAG, which uses full-text-search or hybrid search, then you need to convert to unescaped markdown, since the keywords from the user's query may contain underscores and the required document chunks will not be found.

I suggest adding an option to export_to_markdown() which disable escaping when exporting to Markdown, and also using it with the False value when converting to text. The default value remains True.