A Production RAG Pipeline for PDFs: Relational Parsing, TOC Retrieval, Typed Answers

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

Summary

The article details significant upgrades to a four-brick RAG pipeline (document parsing, question parsing, retrieval, generation) for enterprise use. Building on a minimal RAG baseline, it addresses common failure points like noisy user input and unstructured PDF data. Document parsing now yields a relational set including "line_df", "page_df", "toc_df", and "parsing_summary". Question parsing transforms user queries into a structured "ParsedQuestion" with corrected keywords and inferred answer shapes. Retrieval shifts from simple keyword matching to structured filtering, leveraging the TOC via an LLM router. Generation provides typed answers with citable spans and quality indicators like "confidence" and "context_structured". These upgrades were tested on the 15-page arXiv paper "Attention Is All You Need" using a noisy question.

Key takeaway

For MLOps Engineers building enterprise RAG systems, prioritizing structured data contracts and intelligent parsing across the pipeline is crucial. Upgrading document parsing to relational sets, question parsing to structured briefs, and generation to typed, citable answers will significantly enhance robustness and auditability. This reduces common failure modes like typo-induced misses or unstructured outputs. Implement TOC-aware retrieval and output quality indicators to improve answer reliability and user experience.

Key insights

Robust RAG pipelines require structured data contracts and intelligent parsing across all stages.

Principles

• Retrieval is filtering, not just search.
• Parse questions to structure user intent.
• Generation needs typed, auditable outputs.

Method

Upgrade RAG by enhancing document parsing to relational sets, question parsing to structured briefs, retrieval with TOC-aware filtering, and generation to typed, citable answers.

In practice

• Use LLMs to interpret document TOCs for retrieval.
• Expand keywords with expert vocabulary.
• Return typed answers with per-item evidence spans.

Topics

• RAG Pipeline
• Document Parsing
• Question Parsing
• Information Retrieval
• LLM Generation
• Enterprise AI
• PDF Processing

Code references

• doc-intel/notebooks-vol1

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

Related on AIssential

• Your RAG Pipeline Is Probably Useless. Here’s a Better Alternative — RAG pipelines fail predictably; alternatives like long-context, summarization, structured, or graph-based retrieval offer better solutions.
• 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.
• LLM-Powered Deep Parsing for Industrial Inventory Search — LLM-powered deep parsing extracts structured data from messy industrial text for improved search and deduplication.
• Multi-Format AI Data Prep for Context-Aware RAG Pipelines — Reliable RAG requires structure-aware data preparation to preserve context across diverse file formats.
• Building a Production-Ready RAG Pipeline Workshop — Building production RAG systems requires advanced techniques beyond basic chunking and embedding.
• Retrieval Is Filtering, Not Search: A Mental Model for Enterprise RAG — RAG retrieval is filtering structured document tables, not free-text search, separating anchor detection from context expansion.

Open in AIssential →

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