Human-Imperceptible Retrieval Poisoning Attacks in LLM-Powered Applications