Three Kinds of RAG Corpus, and What It Costs to Build for the Wrong One

Standard Retrieval-Augmented Generation (RAG) systems that treat all documents as a single flat pile fail at enterprise scale due to issues like context fragmentation and vocabulary drift. These problems cannot be solved by simply tuning parameters or adding a re-ranker layer. A more effective approach is to first classify the document collection into one of three shapes: a pile of unrelated files, a database of structured documents, or a set of case bundles. This classification, determined by asking three key business questions, dictates the required system architecture, allowing for structural filtering before vector search, which is more efficient and accurate.