The prevailing approach to Retrieval-Augmented Generation (RAG) often involves a standard playbook: chunking documents, embedding vectors, and retrieving top matches for a Large Language Model. However, this generic strategy frequently fails because it ignores the specific nature of enterprise data. A more effective approach requires diagnosing the problem along two axes: document complexity and question control, then selecting the matching technique. Document complexity ranges from fixed templates to visually rich schematics. Tier 1 includes fixed-structure documents like insurance certificates or tax filings produced by the same software. For these, regex or coordinate-based extraction is sufficient and far more efficient than using an LLM. Tier 2 covers families of templates with minor variations, such as invoices from different suppliers, requiring a mix of regex and lightweight fallback models. Tier 3 involves heterogeneous structured documents like custom contracts, where parsing the table of contents aids retrieval. Tier 4 encompasses unstructured or scanned text, needing optical character recognition and hybrid retrieval. Tier 5 consists of visually rich content like engineering schematics or charts, which demand vision-capable models since text-only parsing fails to capture the meaning. The question axis defines who controls the inquiry. Engineer-templated questions (Tier A) are fixed prompts used for data extraction, while user-filled slots (Tier B) allow for limited variables. Free, one-shot user queries (Tier C) represent the classic chat-with-your-document scenario, and Tier D adds a clarification loop where the system can ask follow-up questions to resolve ambiguity. Cross-referencing these axes reveals distinct technical zones. The top-left corner, involving fixed templates and controlled questions, is deterministic territory. Teams often over-engineer here by deploying LLMs for tasks solvable by regex. The central band, covering heterogeneous documents and open queries, is the domain of full single-document RAG, utilizing chunking, retrieval, reranking, and evaluation. The bottom row is reserved for vision tasks where visual data is primary. Additionally, corpus-scale questions targeting multiple documents require a different stack involving SQL aggregation and structured field extraction rather than pure RAG. Selecting the simplest effective technique is critical. Long context windows do not replace the need for retrieval, and advanced methods like hypothetical document embeddings often mimic keyword search at a higher cost. The most robust production systems use a hybrid approach: a deterministic core for the majority of cases, with LLMs reserved for edge cases where format rules break. Before building, teams must identify their specific use case through a diagnostic process. They should determine who the user is, typically an expert familiar with the domain, and frame the system to amplify that expert rather than replace them. By answering questions about document structure, question control, and constraints, teams can locate their problem on the grid and choose the appropriate technique. Starting with the right diagnostic prevents wasted resources and ensures the solution scales effectively.