The article presents ten positions for Enterprise Document Intelligence that challenge mainstream RAG tutorials. It argues RAG is not machine learning; embeddings are not magic; chunk-size sweep optimizes the wrong thing; answer schema matters more than the model. The standard tutorial pattern—chunk documents, embed, retrieve by cosine similarity, rerank, send to LLM—works on hello-world examples but wobbles with real enterprise documents.

The ten positions fall into three layers: retrieval (structure first, dictionaries before models, rerankers as tools not stages, one vector store for everything), enterprise definition (what enterprise means, who the system amplifies, who picks routes at runtime), and audit (per-failure evaluation, relational structure between bricks, citations as evidence). Position 1 inverts the default tutorial: vector stores are a fallback, not a foundation. Structure-first retrieval handles bulk of real questions; embeddings come in as safety net for residual cases.

The series companion notebooks run on real PDFs, demonstrating structure-first retrieval before embedding, typed answers with line-level citations, evaluation sliced by failure mode.