One RAG Pipeline, Four Very Different PDFs: Same Four Bricks, Every Answer Typed and Cited

2026-07-17 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, extended

Summary

A single, upgraded RAG pipeline ("pdf_qa" function) is presented, designed to process diverse PDF documents without code changes. It builds upon four "bricks": document parsing, question parsing, retrieval, and generation, each previously enhanced. The pipeline was tested on four distinct PDFs: "Attention Is All You Need" (research paper), NIST Cybersecurity Framework 2.0 (compliance document), "Retrieval-Augmented Generation" (research paper), and a World Bank "Commodity Markets Outlook" (report with broken TOC). For the "Attention" paper, it answered a listing question about positional encoding. For NIST CSF 2.0 and the RAG paper, it answered single-concept questions. Crucially, for the World Bank report with a degenerate TOC, the pipeline correctly declined to answer, returning "NA" with 0.0 confidence, demonstrating robustness against hallucination when context is insufficient. The pipeline incorporates two side channels: "parsing_summary" feeding question parsing, and a feedback loop from generation to retrieval. The article emphasizes the independence and typed contracts of each brick, allowing for modular upgrades.

Key takeaway

For AI Engineers building enterprise RAG systems, you should prioritize a modular, "brick-by-brick" architecture with explicit typed contracts. This approach ensures your pipeline can adapt to diverse document types and questions without code changes, significantly reducing maintenance. Implement robust context engineering, including TOC-aware retrieval and generation feedback, to prevent hallucinations and provide auditable, cited answers, even when source context is incomplete.

Key insights

A robust RAG pipeline uses independent, typed "bricks" and context engineering to handle diverse documents and prevent hallucinations.

Principles

• Decompose RAG into independent, typed bricks.
• Explicit contracts enable modular upgrades.
• Context engineering prevents LLM hallucinations.

Method

The "pdf_qa" function orchestrates four bricks: parsing (relational data), question parsing (structured brief), retrieval (keyword + TOC routing), and generation (typed, cited answer with quality indicators). It includes "parsing_summary" and a generation-to-retrieval feedback loop.

In practice

• Implement typed contracts between RAG components.
• Use TOC-aware retrieval for long documents.
• Return confidence scores and completeness flags.

Topics

• RAG Pipeline Architecture
• Document Intelligence
• PDF Parsing
• Context Engineering
• LLM Hallucination Prevention
• NIST Cybersecurity Framework

Code references

• doc-intel/notebooks-vol1

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Your PDF Isn’t Just Text: Building a Production-Ready Multimodal RAG Pipeline — Preserve document structure and relationships throughout the RAG pipeline to prevent context loss and hallucinations.
• Context Engineering for RAG : The Four Typed Inputs Behind Every RAG Answer — Context engineering systematically assembles typed inputs for LLMs, enhancing RAG auditability and cost-efficiency.
• Enterprise Document Intelligence: A Series on Building RAG Brick by Brick, from Minimal to Corpus scale — Enterprise RAG success hinges on deep document and domain expertise, not just advanced AI models.
• Context Engineering for Bedrock Agents: A Hands-On Guide Beyond Prompt Engineering — Layered context pipelines are crucial for preventing hallucination in multi-turn, tool-using LLM agents.
• LAI #121: The single-agent sweet spot nobody wants to admit — Effective AI system design requires distinguishing agents from workflows and evaluating RAG pipelines in two distinct layers.
• Your RAG Pipeline Isn’t Broken. Your Chunks Are. — Bad document chunking silently destroys RAG pipeline performance, often misattributed as LLM hallucination.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards Data Science.