Stop Returning Flat Text from a PDF: The Relational Tables RAG Needs

2026-06-11 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Intermediate, extended

Summary

This article describes a PDF parsing approach for RAG systems that models documents as a relational set of tables instead of flat text. It details eight specific DataFrames: toc_df, page_df, line_df, image_df, span_df, object_registry, cross_ref_df, and parsing_summary. These tables capture document structure, content, typography, images, and cross-references, enabling downstream RAG components (retrieval, generation, highlighting) to query structured data. The parse_pdf function generates these linked tables, demonstrated on a 15-page LaTeX research paper ("Attention Is All You Need") and a 32-page NIST Cybersecurity Framework 2.0 document, showing consistent output structure. This method replaces costly re-parsing with efficient DataFrame queries, improving RAG pipeline performance and accuracy, especially for complex documents like contracts with tables.

Key takeaway

For AI Engineers building RAG pipelines for complex enterprise documents, you must move beyond simple text extraction. Implement a relational parsing approach that generates linked DataFrames like line_df and toc_df. This preserves critical structural and semantic information, preventing issues like lost table context. By caching these parsed tables, you significantly reduce re-parsing costs and enable more accurate, context-aware retrieval and generation, improving overall RAG system performance.

Key insights

PDF parsing for RAG should produce a relational set of linked tables, not flat text, to preserve document structure.

Principles

• Model documents as relational tables.
• line_df is the central source of truth.
• Semantic signals improve retrieval.

Method

The parse_pdf function generates eight linked DataFrames (toc_df, page_df, line_df, image_df, span_df, object_registry, cross_ref_df, parsing_summary) from a PDF, enabling structured queries.

In practice

• Use toc_df for section-level retrieval.
• Filter line_df by column_position for tables.
• Enrich image_df with vision LLM descriptions.

Topics

• RAG Systems
• PDF Parsing
• Document Intelligence
• Relational Data Models
• Information Extraction
• DataFrames

Best for: AI Engineer, MLOps Engineer, Data Engineer

Related on AIssential

• Beyond extract_text: The Two Layers of a PDF That Drive RAG Quality — Effective RAG parsing requires understanding both PDF structural signals and page content, augmented by an LLM-generated semantic summary.
• Your RAG Isn’t Broken. Your Table Headers Are. — Stripped table headers silently destroy RAG retrieval quality by removing essential contextual schema from data.
• Parse PDFs for RAG Locally with Docling: Rich Tables, No Cloud Upload — Docling offers robust, local PDF parsing for RAG, overcoming cloud compliance and cost issues with rich table and layout extraction.
• Article: Redesigning Banking PDF Table Extraction: A Layered Approach with Java — Reliable PDF table extraction in finance requires a layered architectural approach, not just a single parsing library.
• Why Naive Chunking Breaks RAG, and What to Build Instead — Structure-aware parsing and multimodal retrieval techniques are crucial for effective RAG on complex documents.
• Your RAG System Isn’t Retrieving. It’s Guessing. — Current Retrieval Augmented Generation (RAG) systems largely rely on vector databases and the assumption that semantic similarity equates to relevance, a principle powering billions in infrastructure.

Open in AIssential →

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