#1★ TOP PICK
Docling
IBM's document converter. Layout-aware PDF to clean Markdown, MIT.
94
OPEN SOURCEMITSELF-HOSTLOCAL-FIRST
Docling parses PDFs, Office documents, images and HTML into a structured representation that preserves reading order, tables, figures and headings, then exports to Markdown or JSON. It uses purpose-trained layout and table models rather than heuristics, which is why it holds up on multi-column academic papers and financial statements where simpler extractors interleave columns into nonsense. It integrates directly with LlamaIndex and Haystack, is MIT licensed, and runs entirely locally including on CPU.
Strengths
- +Layout-aware — preserves reading order, tables and structure
- +Outputs clean Markdown/JSON that drops into a RAG pipeline
- +Direct integrations with LlamaIndex and Haystack
- +MIT, fully local, no per-page cost
Trade-offs
- −Slower per page than cloud OCR on very large batches
- −Handwriting support is weak compared with Textract
- −No specialised invoice or receipt models
Free, MIT. Runs on your own hardware, CPU or GPU. #2
Tesseract
Thirty years old, Apache-2.0, and it runs absolutely everywhere.
95
OPEN SOURCEApache-2.0SELF-HOSTLOCAL-FIRST
Tesseract is the OCR engine most of the world's text extraction has quietly gone through, originally from HP, developed by Google for a decade, and Apache-2.0 throughout. It recognises over a hundred languages, has bindings for every language you might be writing in, and is packaged in every Linux distribution. Modern neural extractors beat it on complex layouts, but on clean scanned text it is fast, dependable and effectively free of dependencies — often the right tool precisely because it is boring.
Strengths
- +Over a hundred languages, bindings for everything
- +Extremely mature and stable — decades of production use
- +Packaged everywhere; trivial to install and deploy
- +Apache-2.0 with no dependency weight
Trade-offs
- −Weak on complex layouts, multi-column pages and tables
- −Needs image pre-processing for good results on poor scans
- −No document understanding — it gives you text, not structure
#3
PaddleOCR
The strongest open OCR engine, and the best at non-Latin scripts.
92
OPEN SOURCEApache-2.0SELF-HOSTLOCAL-FIRST
PaddleOCR is a comprehensive OCR toolkit covering text detection, recognition, table extraction, layout analysis and key-value extraction, with support for around eighty languages. Its non-Latin script handling — Chinese, Japanese, Korean, Arabic — is clearly the best of the open options, and it ships lightweight mobile-scale models alongside the accurate server ones. Apache-2.0. If raw recognition accuracy on difficult scans is the constraint, this is the engine.
Strengths
- +Best-in-class open recognition accuracy on difficult scans
- +Around eighty languages, with excellent non-Latin coverage
- +Includes table, layout and key-value extraction
- +Lightweight models for edge and mobile deployment
Trade-offs
- −Built on the PaddlePaddle framework — an extra dependency to adopt
- −Documentation is stronger in Chinese than in English
- −Output needs more post-processing than Docling's Markdown
#4
Unstructured
One interface for every document format you will actually be handed.
90
OPEN SOURCEApache-2.0SELF-HOSTLOCAL-FIRST
Unstructured normalises an unusually wide range of inputs — PDF, Word, PowerPoint, Excel, HTML, email, EPUB, images — into a consistent element structure of titles, narrative text, tables and lists, ready for chunking. Breadth is the point: real corpora are never one format, and writing a parser per type is where ingestion projects stall. The open library is Apache-2.0 and runs locally; the company also sells a hosted API with additional models.
Strengths
- +Widest input-format coverage of anything here
- +Consistent element output regardless of source format
- +Chunking strategies built in for RAG pipelines
- +Apache-2.0 open library
Trade-offs
- −Best-quality models sit in the paid hosted tier
- −Local install pulls in heavy system dependencies
- −Depth on complex PDFs is below Docling's
Open library free under Apache-2.0; a paid hosted API is offered separately. #5
Marker
The best PDF-to-Markdown quality here — check the licence first.
70
SOURCE-AVAILABLEGPL-3.0 with commercial revenue conditionSELF-HOSTLOCAL-FIRST
Marker converts PDFs to Markdown with the highest fidelity of anything in this list: it handles equations, code blocks, tables and multi-column layouts, and optionally uses an LLM pass to improve difficult sections. For academic papers and technical documents the output quality is noticeably ahead. The caveat we would rather state than have you find in a licence review: Marker is GPL-3.0 with an additional revenue condition from its maintainer, so organisations above a revenue threshold need a commercial licence. Free for research and for smaller organisations, but not unconditionally.
Strengths
- +Best PDF-to-Markdown fidelity of the options here
- +Handles equations, code blocks and complex tables well
- +Optional LLM pass for difficult pages
- +Runs entirely locally
Trade-offs
- −Not unconditionally open source — revenue-gated commercial terms
- −GPL-3.0 copyleft affects how you can distribute derived work
- −GPU strongly recommended for reasonable throughput
Free below the maintainer's revenue threshold; a commercial licence is required above it.