Grounding is a strategy used to enhance the pre-trained knowledge of Large Language Models (LLMs) by incorporating relevant external information along with the task prompt, with retrieval-augmented generation (RAG) being the leading method. The process addresses inherent knowledge gaps in LLMs that arise from their limited training data and finite parameters, by providing additional context that helps reduce the likelihood of hallucinations, where LLMs generate plausible but incorrect information. Effective grounding involves ensuring data relevance, quantity, and arrangement, with challenges like interpreting query intent and mitigating the "lost in the middle" bias, which affects how LLMs process information. The effectiveness of grounding is influenced by the quality of external data, and ongoing research focuses on improving provenance and the ability to update models post-training, thereby enhancing the coverage of pre-trained LLMs and reducing reliance on external information.