Extracting structured information from visually rich documents like invoices and receipts is a complex task for AI. Traditional methods often struggle to generalize to new, unseen document types, while leveraging powerful large language models (LLMs) like GPT-4 presents unique challenges. How do you teach an LLM to understand the intricate relationship between textual elements and their layout on a page? Researchers are tackling this problem with innovative techniques. One exciting new approach called SAIL (SAmple-centric In-context Learning) focuses on providing LLMs with more targeted learning experiences. Instead of simply feeding the model a few generic examples, SAIL crafts custom-tailored prompts for each document. It considers the layout similarity by analyzing the visual structure of the documents and finding similar examples. It also dives deeper into the text, identifying entities with similar meanings to help the LLM grasp the nuances of the content. By presenting a variety of relevant examples, SAIL guides the LLM to extract the correct information with remarkable accuracy, even surpassing existing methods in some cases. The results are promising, demonstrating a significant improvement in information extraction performance across various LLM models. This sample-centric approach opens up new possibilities for training-free document processing, suggesting a future where AI can effortlessly extract data from any document, regardless of its format. While challenges remain, including the computational cost of searching for similar examples and the limits on prompt length, SAIL represents a significant step forward in unlocking the wealth of information trapped within our documents.