Comprehensive Benchmarking of Python Text Extraction Libraries: Insights from 2025 Results
In the pursuit of optimized text extraction in Python, selecting the right library can be a game-changer—especially when dealing with diverse document types and sizes. Recently, I undertook an extensive benchmarking project in 2025 to evaluate four prominent Python text extraction libraries, providing clarity on their performance, reliability, and suitability for various use cases.
Why Benchmark?
As the creator of Kreuzberg, a library designed for efficient text extraction, I sought an objective, data-driven comparison to understand how Kreuzberg stacks up against other established tools. This benchmarking was performed on a broad spectrum of real-world documents, ensuring the results reflect practical scenarios rather than synthetic tests.
The Competitors
The evaluation covered the following libraries:
-
Kreuzberg
A lightweight, fast, and versatile Python library with an easy-to-use API, supporting both synchronous and asynchronous operations with OCR capabilities. -
Docling
An enterprise-focused solution powered by machine learning, notable for its advanced understanding of complex documents but with significant resource requirements. -
MarkItDown
A Microsoft-developed Markdown conversion tool optimized for straightforward PDFs and Office documents, suitable for preprocessing before large language model (LLM) tasks. -
Unstructured
An open-source framework aimed at processing diverse enterprise documents with a focus on reliability and support for multiple languages.
Testing Scope
The benchmarking involved analyzing 94 real-world documents spanning formats like PDFs, Word files, HTML, images, and spreadsheets. These documents ranged from tiny files under 100KB to massive academic papers exceeding 50MB, covering multiple languages such as English, Hebrew, German, Chinese, Japanese, and Korean. To maintain fairness, all tests were conducted on CPUs without GPU acceleration, measuring metrics including:
- Processing speed
- Memory consumption
- Success rates
- Installation size
Key Findings
Performance Highlights
| Rank | Library | Average throughput | Remarks |
|———|————-|————————-|—————————————–|
| 1 | Kreuzberg | 35+ files/sec | Consistently rapid and robust |
| 2 | Unstructured | Moderate speed | Excellent reliability, handles complexity well |
| 3 | MarkItDown | Good on simple files | Struggles with complex documents |
| 4 | Docling | Up
