A production RAG pipeline for real-world PDFs: structural retrieval, typed answers, cited lines

Last Updated on July 6, 2026 by Editorial Team Author(s): Angela Shi Originally published on Towards AI. The four bricks, run end-to-end on a real 45-page car-insurance policy. One surprising coverage question, answered with a number and the exact line it came from Use this link if you are not a member. Page 30 as the pipeline sees it: every line boxed, its number beside it. The pet-injury block wraps from the left column into the right; the answer is on lines 54 and 55. The pipeline routes here by section name, then cites those exact lines. — image by authorAfter the lead, the article walks through a production-ready RAG approach built as four end-to-end “bricks” rather than a naive embedding-and-generation baseline: it parses real PDFs into relational tables (lines, bounding boxes, pages, and a reconstructed table of contents), turns the user question into a typed brief that includes intent, expected answer shape (e.g., capped dollar amount), section hints, and expert vocabulary, routes retrieval by section filtering using an anchor-and-router pattern to avoid confusing similarly embedded caps, and finally generates a constrained “contract” output (typed value, evidence span with exact line coordinates, confidence, and caveats). It concludes by showing how the same typed pipeline scales to harder question types like listings, decompositions, and scoped synthesis, emphasizing that auditable intermediate objects and a consistent contract make it reliable for real-world document QA. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI