TCRA-LLM: Token Compression Retrieval Augmented Large Language Model for Inference Cost Reduction F i n d i n g s - E M N L P