Reduce LLM hallucination and handle class imbalance

Answer the following applied ML/LLM questions. ## 1) LLM hallucination & token cost control You are building a chatbot over an internal knowledge base. 1. What are common causes of hallucination in LLM chatbots? 2. How would you reduce hallucinations using a RAG-style approach (retrieval + generation)? 3. How would you control or reduce token costs while maintaining answer quality? Discuss concrete design choices (e.g., chunking, retrieval quality, prompt construction), evaluation ideas, and failure modes. --- ## 2) Class imbalance + precision vs. recall trade-off You are building a binary classifier where the positive class is rare. 1. How would you handle class imbalance during training and evaluation? 2. In a scenario where **false positives are more costly than false negatives**, which metric should be prioritized (precision vs. recall), and how would you choose an operating threshold? Explain your reasoning and mention practical checks/pitfalls (e.g., calibration, dataset shift).