Hallucinations are the core reliability challenge for LLM applications. Root causes: 1) Training data limitations (outdated, incomplete, inaccurate). 2) Over-confident generation (model continues fluently even when uncertain). 3) Instruction following trying to answer questions it should decline. Proven techniques: 1) RAG with faithfulness constraint: instruct model "Only use information from the provided context. If the answer is not in the context, say I don't know." Reduces hallucination significantly for factual queries. 2) Self-consistency: sample 5-10 responses with temperature > 0, take majority vote or detect agreement. Aggregated answer is more reliable than single sample. 3) Chain-of-verification: generate answer, generate verification questions about specific claims, check each claim independently, revise answer if verification fails. 4) Retrieval verification: after generating response, extract claims, verify each against retrieved documents, provide confidence score. 5) Calibration prompt: add "If you are not sure, say 'I am not certain but...'" - increases appropriate hedging. 6) Constitutional self-critique: have model review its own output for hallucinations before returning. 7) Fine-tuning for calibration: include uncertainty examples in training data so model learns to hedge appropriately. Evaluation: FactScore (fine-grained factuality evaluation) and FELM measure hallucination rates. Typical production RAG achieves 80-90% faithfulness; techniques above improve to 92-97%.