Hello, I am seeking guidance on using RAG to ground a model response format based on DSL semantics unknown by the LLM. My task is to assess how RAG can help align model responses with a specific JSON schema, which can be complex. As I am new to building RAG-augmented LLM apps, any advice on evaluating efficiency before starting the entire pipeline would be greatly appreciated. I can't rely on function callings nor newly introduced OpenAI feature (JSON mode, seeds sampling). I am trying to base my approach on the findings of this paper: https://arxiv.org/pdf/2308.00675.pdf where they seem to indicate you can teach an LLM to use a new tool/language by passing the documentation as input