RAG与长上下文：如何向大型语言模型提供私有数据？

Large Language Models (LLMs) are fixed in time, knowing only what they were trained on up to their cutoff date and lacking access to internal documents unless context is injected at query time. Two engineering patterns compete: Retrieval-Augmented Generation (RAG), which involves chunking documents, embedding, storing vectors, retrieving top matches, and injecting snippets into prompts; and Long-context (brute force), which dumps large documents directly into the model's context window, allowing attention to find answers. This architectural choice impacts complexity, reliability, cost, and correctness. Choosing the wrong pattern can lead to either (a) missing facts due to retrieval failures or (b) exceeding budget while still obtaining noisy results because the model cannot focus on the critical information amidst the clutter.