clean up redundancies in prompts and ensure tokens are being reserved for generation when filling semantic context