Why PDF Parsing Is Hard for RAG and Which Mainstream Solutions Work

Document Format Design

DOCX stores explicit semantic structure in XML, HTML uses semantic tags and CSS for presentation, while PDF is a page description language that records only coordinates, glyphs, lines, colors, and drawing operators without inherent paragraphs, headings, or tables.

PDF Structure

A PDF consists of Header, Body, Cross‑Reference Table, and Trailer. The Body contains COS objects such as Dictionary (describes structure), Stream (actual content data), Array (page tree, font list), String (text), and Indirect Reference (e.g., 50R refers to object 5).

The xref table indexes each object’s byte offset, enabling random access without scanning the whole file. Text resides in each page’s content stream, with fonts and images referenced via a Resource Dictionary.

Structural Parsing Difficulties

No explicit reading order : Characters are stored in rendering order, which may differ from logical order, especially in multi‑column layouts, making it hard to reconstruct correct flow.

Tables lack data structure : Tables are visual line‑based constructs; parsers must infer cell boundaries from intersecting lines, often failing on merged cells.

Font‑encoding mapping : PDFs store glyph IDs, not Unicode. Missing or incomplete ToUnicode CMaps cause garbled output or missing characters.

Scanned PDFs have no text layer : They contain only raster images, requiring OCR before any textual extraction.

Parsing Principles: From Byte Stream to Readable Text

Native Text Extraction

This fastest route works on PDFs generated directly from Word, LaTeX, InDesign, etc., which embed a complete text layer.

Main workflow:

BT                     % Begin Text object
 /F1 12 Tf             % select font F1, size 12pt
 72 720 Td             % move to (72,720)
 (Hello, PDF!) Tj      % show string
 0 -14 Td              % move to next line
 [(Rag) 20 ( pipeline)] TJ % TJ operator
ET                     % End Text object

The parser reads each operator, maintains a graphics state machine (CTM, current font, text matrix) and finally obtains absolute character coordinates. The difficulty lies in aggregating discrete glyph points into words, sentences, and paragraphs; libraries such as PyMuPDF and pdfplumber differ mainly in their aggregation algorithms.

OCR

For image‑only PDFs, the page is rasterized and fed to an OCR engine.

Tesseract : Performs connected‑component analysis then LSTM‑based transcription; assumes single‑column layout and can be unstable on complex pages.

PaddleOCR : Adds a layout‑analysis model to detect regions before OCR, improving multi‑region handling.

Challenges include table reconstruction after OCR and cascading errors from inaccurate layout detection.

Visual Language Models (VLM)

VLMs treat the whole page as an image and output structured text, bypassing coordinate extraction, font decoding, and reading‑order reconstruction.

Advantages: natural handling of multi‑column, nested tables, mixed text‑image layouts; understands merged cells and cross‑page tables.

Drawbacks: hallucinations, high computational cost, and potential data‑privacy concerns.

Mainstream Parsing Solutions

Lightweight text‑extraction libraries : PyMuPDF, pdfplumber, pypdf – extremely fast but limited by PDF’s inherent structural gaps.

AI‑enhanced open‑source frameworks : Docling, MinerU, Marker‑PDF – integrate layout‑analysis models (e.g., DocLayNet, PP‑StructureV2) and specialized table‑reconstruction modules.

Commercial APIs : Cloud providers offer hosted PDF‑parsing services that are ready‑to‑use.

Direct VLM calls : GPT‑4o, Gemini 2.5, Qwen3‑VL – deliver the strongest understanding at the highest cost.

Lightweight Libraries

These tools read character coordinates and sort them, without any layout analysis. They excel on well‑structured PDFs but struggle with tables lacking explicit borders or with merged cells.

Examples:

PyMuPDF : Built on MuPDF, high compatibility, stable accuracy on clean documents.

PyMuPDF4LLM : Extends PyMuPDF for LLM use, adding Markdown formatting, simple table reconstruction, and heading detection.

pdfplumber : Based on pdfminer.six, focuses on table extraction by detecting horizontal and vertical lines; works well on clearly bounded tables but poorly on border‑less or merged‑cell tables.

AI‑Enhanced Frameworks

These add a middle layer that performs region classification (title, paragraph, table, image) using models such as DocLayNet, then reconstructs a document tree (e.g., DoclingDocument) preserving hierarchy.

Docling can export to Markdown, HTML, JSON, or DocTags, enabling downstream chunking based on semantic structure rather than raw text.

MinerU leverages PaddleOCR for OCR and a PDF‑Extract‑Kit layout detector, showing strong results on Chinese documents and LaTeX papers.

Marker‑PDF provides a unified entry point for many formats (PDF, DOCX, PPTX, etc.), simplifying heterogeneous document pipelines.

Commercial APIs

Major cloud vendors offer hosted PDF‑parsing APIs; the article does not analyze them in depth.

VLM Direct Calls

VLMs rasterize pages and feed them to multimodal models, achieving the most comprehensive understanding of complex layouts, merged cells, and handwritten content.

Common Parsing Problems and Remedies

Multi‑column Layouts

Characters from left and right columns interleave in the content stream, causing mixed‑order output. Simple coordinate‑based re‑sorting (by y then x) helps for two‑column cases but fails on three‑plus columns or mixed text‑image layouts.

Best practice: run a layout‑detection model (e.g., DocLayNet, PP‑StructureV2) to segment the page into independent regions, assign reading order, and concatenate region text accordingly. For extremely complex layouts, resort to VLMs.

Font‑Encoding Gaps

Missing or incomplete ToUnicode CMaps lead to garbled characters. Different parsers handle this variably; switching from pdfplumber to PyMuPDF (or vice‑versa) can sometimes resolve the issue.

If all parsers fail, OCR the page image to bypass font‑encoding entirely.

PDF 中显示：Attention Is All You Need
PyPDF2 提取：偛整匯數整數搔搔数搔數

Cross‑Page Content Breaks

Physical pagination does not align with logical segmentation, causing paragraphs, sentences, tables, figures, or code blocks to split across pages.

Solutions:

Introduce overlapping tokens between adjacent chunks to keep split sentences intact.

Use semantic chunking based on embeddings to cut at true semantic boundaries.

Detect unfinished tables at page end and merge with the next page’s table rows (supported by Docling and MinerU).

Apply layout analysis to re‑assemble multi‑page structures.

Header/Footer Pollution

Repeated headers/footers generate near‑duplicate chunks, inflating similarity scores and harming retrieval.

Typical fix: filter out text whose y‑coordinate falls within the top/bottom 8 % of the page, then perform cross‑page duplicate detection by counting line frequency across pages and discarding lines exceeding a configurable threshold.

Loss of Figures and Charts

Pure text extraction drops embedded images, charts, and diagrams, leaving empty references.

Remedy: combine text extraction with OCR to recover any embedded textual annotations, then pass the rasterized page to a VLM for visual understanding of the chart’s semantics.

References and Footnotes

References and footnotes are often far from their citation markers, causing them to be placed in unrelated chunks.

Layout models can separate footnote regions; frameworks like Docling and MinerU expose footnotes as distinct nodes that can be attached to the citing paragraph or stored separately with metadata for later retrieval.

For bibliography entries, embed the full reference text into the metadata of the citing chunk or, after retrieval, fetch the corresponding bibliography item and let the LLM combine both.

Lost Heading Hierarchy

Fixed‑size token chunking discards document hierarchy, making it impossible to know a chunk’s level.

Simple fix: attach the full hierarchical path (e.g., "Section 2 > Subsection 2.1") as metadata to each chunk. A more advanced approach is to chunk by semantic boundaries (sections, subsections) rather than token count, preserving structural integrity.

Conclusion

PDF parsing lacks a perfect solution because the format’s design omits semantic information. Heuristic methods, statistical models, and increasingly powerful AI (layout analysis, VLMs) are used to reconstruct the missing semantics. For demanding production scenarios, a hybrid approach—combining lightweight extraction, AI‑enhanced layout analysis, and VLMs—offers the best trade‑off, albeit with added complexity and cost.