One option would be similar to function calling, give the llm an output it can make that changes how the context is parsed. That's a layer on top rather than changing how the llm itself works.