The article maps enterprise retrieval‑augmented generation (RAG) onto a two‑axis diagnostic that separates document structure from question control. It argues the ubiquitous classic RAG playbook—chunking PDFs, embedding, vector search, then feeding results to an LLM—covers only a middle band of use cases. Templated forms, sarcastic call transcripts, and image‑heavy schematics each demand a different stack. It illustrates extremes: high‑volume templated forms, sarcasm‑laden transcripts, and pixel‑level schematics.

Three document tiers illustrate the mismatch. Tier 1 fixed‑template PDFs—insurance certificates, KYC forms, payroll stubs—are cheap to parse with regex or coordinate extraction, avoiding LLM cost. Tier 4 unstructured scans and Tier 5 visually rich pages require OCR‑guided hybrid retrieval or vision models that read charts, diagrams, and handwritten notes. Choosing the wrong stack can inflate per‑document processing by orders of magnitude.

The diagnostic proposes three steps: locate your position on the structure‑vs‑control grid, match the region to its recommended stack, then apply the appropriate agentic layer that governs LLM runtime. Most deployments collapse into field extraction from templates or free‑form Q&A over heterogeneous contracts. By forcing teams to diagnose first, the guide prevents over‑engineering and unnecessary cloud spend, steering projects toward the most economical technique.