Large language models are increasingly being introduced into pharmaceutical knowledge management to support regulatory intelligence, safety surveillance, scientific literature review, and document search. Retrieval-augmented generation has become especially attractive because it promises to ground responses in approved labels, scientific publications, patents, and internal reports. Despite this enthusiasm, the evidence base remains uneven and fragmented. Current systems often demonstrate impressive linguistic fluency, yet fluency is frequently mistaken for factual reliability, domain competence, and regulatory readiness. This critical review evaluates the use of large language models and retrieval-augmented generation in pharmaceutical knowledge management. It focuses on hallucination control, domain-specific evaluation, trustworthiness, and the conditions required for safe deployment in regulated workflows. Retrieval-augmented generation reduces some factual errors but does not eliminate hallucination, source misuse, or incomplete reasoning. Evaluation methods remain immature, with many studies relying on metrics that do not adequately measure pharmaceutical correctness, completeness, or actionability. Unverified outputs from large language models may create risks for patient safety, pharmacovigilance, regulatory compliance, and internal decision-making. Responsible deployment requires expert oversight, traceable sources, robust evaluation, and explicit governance rather than confidence in model fluency alone. Large language models may become valuable tools for pharmaceutical knowledge work, but they should not yet be treated as autonomous knowledge authorities. The field must build pharmaceutical-specific benchmarks, stronger fact-checking protocols, and auditable governance frameworks before these systems can be trusted in high-stakes contexts.