Adverse drug events (ADEs) contribute significantly to hospital admissions and increased mortality rates. Early recognition of these reactions remains crucial. Traditional systems like the FDA Adverse Event Reporting System (FAERS) have been key to drug safety surveillance, but social media platforms like Reddit now offer a rich source of real-world insights, transforming pharmacovigilance. To tackle the challenge of extracting meaningful insights from vast, unstructured data, Large Language Models (LLMs) offer an invaluable solution. In this paper, we present “In-Context Pruning Framework” (ICPF), a novel methodology that utilizes a fine-tuned LLaMA-3 model to filter irrelevant content and focus on context-sensitive ADE analysis. Our approach combines SciBERT for entity extraction, and synonym normalization to enhance data consistency. We demonstrate the applicability of this methodology in the domain of Psychiatry by identifying ADRs associated with Psychiatric drugs. Additionally, we conduct co-occurrence and drug class-level analysis to uncover intricate relationships between drugs and their adverse effects. Discrepancy between the ADE extracted from Reddit and the official source is also being presented thereby bridging the gap between user-reported experiences and official sources, providing actionable insights to improve pharmacovigilance and patient safety.