Baseline Enterprise RAG, From PDF to Highlighted Answer

A hands-on walkthrough of a minimal enterprise RAG pipeline built in ~100 lines of Python with no vector database or framework. The pipeline has four bricks: document parsing (PDF to line-level DataFrame with bounding boxes via PyMuPDF), question parsing (LLM-extracted keywords), retrieval (keyword matching vs. embeddings with a detailed comparison of their failure modes), and generation (structured JSON answer with page/line citations via a Pydantic schema). The final step optionally annotates the source PDF by drawing rectangles around cited lines. The article uses the 'Attention Is All You Need' paper and a World Bank commodity report as test documents, demonstrating correct answers, clean 'not found' handling, and honest discussion of where each component breaks. It serves as the first installment of a longer series on enterprise document intelligence.