7 open-source tools that turn PDFs into LLM-ready text

Copy-paste from a PDF gives you broken columns, lost tables and missing text, and your LLM inherits the mess. These seven open-source tools turn PDFs, scans and Office files into clean Markdown and JSON instead.

Every pick is open source (one AGPL, two with extra license terms, all flagged), pushed a commit in the last 90 days (checked 2026-11-11), and installable from its own docs. Ordered as a pipeline: quick dumps first, deep parsing next, raw OCR engines last.

The 7 tools at a glance #

1. MarkItDown: everything into the context window #

MarkItDown is Microsoft's lightweight Python utility that converts files to Markdown for LLMs: PDF, PowerPoint, Word, Excel, HTML, EPubs, ZIP archives, CSV/JSON/XML, even YouTube URLs and audio transcription.

What: One convert() call per file — built for feeding chat context, not for pixel-perfect human documents.

Why: One dependency covers most office formats and keeps headings, lists, tables and links.

Use when: You want a document's content in the conversation in seconds.

Install:

pip install "markitdown[all]"

Catch: Image captioning and speech transcription rely on web-hosted services — not everything stays on your machine.

2. PyMuPDF: CPU-only speed, no GPU #

PyMuPDF is a PDF library that now outputs LLM-ready Markdown and JSON itself (doc.to_markdown()), without a GPU. It is the fast route when you have thousands of text PDFs.

What: A Python library with Markdown output and table extraction built in.

Why: No accelerator needed: it parses text PDFs fast on a plain CPU.

Use when: Batch-converting digital-born PDFs where speed beats layout intelligence.

Install:

pip install --upgrade pymupdf

Catch: AGPL-3.0: closed-source products need Artifex's commercial license instead.

3. Docling: the deep reader #

Docling is a document-parsing toolkit started at IBM. It reads page layout, reading order, table structure, code and formulas, and turns bar and pie charts into tables.

What: A CLI and Python API that plug into LlamaIndex and LangChain pipelines.

Why: It reads structure — which cells belong to which table, what order paragraphs go in — not just text.

Use when: RAG quality matters more than raw speed.

Install:

pip install docling

Catch: It runs layout and table models locally, so it is slower than plain text extractors, and the first run downloads weights.

4. marker: accuracy-first PDF to Markdown #

marker converts PDFs to Markdown and JSON with layout models, and can call an LLM to fix tables and math.

What: A pip package and CLI for demanding, mixed-layout documents.

Why: On olmocr-bench, a third-party test of 1,403 PDFs, its README reports 76.0%, ahead of MinerU and Docling.

Use when: Extraction quality is the product, and you can live with model-weight terms.

Install:

pip install marker-pdf

Catch: The code is Apache-2.0, but the model weights use a modified OpenRAIL-M license — free for research, personal use, and startups under $5M funding or revenue.

5. MinerU: complex documents, agent-ready #

MinerU turns complex PDFs and Office files into LLM-ready Markdown and JSON. It also ships a skill, so coding agents such as Claude Code can run it for you.

What: A converter tuned for hard layouts: scientific papers, multi-column text, formulas and tables.

Why: Where simple extractors lose the structure, MinerU keeps it.

Use when: Your documents are the messy kind — papers, reports, filings.

Install:

pip install uv
uv pip install -U "mineru>=4.0,<5"

Catch: Apache-2.0 with extra terms: online services must credit MinerU, and a commercial license is required above 100 million monthly active users or USD 20 million monthly revenue.

6. PaddleOCR: the multilingual engine #

PaddleOCR is a full OCR (text recognition from images) toolkit for 100+ languages, with Markdown and JSON output. Its small PaddleOCR-VL-1.6 model scores 96.3% on the OmniDocBench v1.6 benchmark.

What: OCR plus document parsing that recovers text, formulas and tables from scans into structured output.

Why: The widest language coverage on this list, with top benchmark scores on formulas and tables.

Use when: The scans are ugly, multilingual, or formula-heavy.

Install:

pip install paddleocr

Catch: It runs on Baidu's PaddlePaddle framework, not PyTorch, so expect a second deep-learning stack.

7. Tesseract: the classic that runs everywhere #

Tesseract is the classic open-source OCR engine. It reads more than 100 languages out of the box and has pre-built packages for every major OS.

What: A battle-tested OCR engine you can call from the CLI, C++ API, or wrappers in every language.

Why: Every major package manager has it, and it needs no GPU or Python.

Use when: You need plain text from images inside your own tooling.

Install:

brew install tesseract

Catch: Outputs plain text, hOCR, PDF, TSV or ALTO, but no Markdown or table structure, so LLM use needs post-processing.

Which one to try first #

Want one file in the chat right now → #1. Thousands of digital PDFs on a CPU box → #2. Structure-aware RAG pipeline → #3. Best Markdown quality → #4. Messy scientific layouts, or driving it from Claude itself → #5. Ugly multilingual scans → #6. Text from images inside your own scripts → #7.