This article details an enhanced Retrieval Augmented Generation (RAG) pipeline designed for processing PDF documents in production environments. The upgraded system addresses limitations found in simpler RAG implementations, particularly when dealing with complex enterprise documents and varied user queries. The pipeline is structured into four core components: document parsing, question parsing, retrieval, and generation, with each element receiving significant improvements.

The document parsing brick now produces a richer, relational dataset beyond just lines of text. This includes a table of contents (TOC), page-level metadata, and a summary of the document's type and content. This structured output aids downstream processes. Question parsing is refined to handle noisy or misspelled user input, correcting keywords and inferring the desired answer format (e.g., single value, list, table).

Retrieval leverages the parsed TOC, using a small LLM to semantically identify relevant sections, combined with keyword matching. The generation brick provides typed answers with verifiable citations and quality indicators. The system is tested on the "Attention Is All You Need" paper, demonstrating its ability to produce structured, citable answers from complex queries, unlike baseline RAG systems that falter with typos or lack of document structure.