Comprehensive Benchmarking of Python Text Extraction Libraries in 2025: Insights for Developers and Data Scientists

In the rapidly evolving landscape of data extraction, choosing the right Python library can significantly impact the efficiency and reliability of your project. Recently, I embarked on an extensive benchmarking journey to evaluate four prominent text extraction tools—aiming to provide clarity and guidance for developers working with diverse document formats and sizes.

Why Benchmark Python Text Extraction Libraries?

With the proliferation of document types—from PDFs and Word files to HTML and scanned images—selecting a performant and reliable library is crucial. While many tools promise impressive features, real-world performance often varies dramatically based on document complexity, size, and language.

To ensure an unbiased and thorough comparison, I tested libraries across a large and varied dataset, measured key performance metrics, and openly share the results to aid your decision-making process.

The Libraries Under Test

The benchmarking focused on these four libraries, representing a spectrum of approaches and capabilities:

• Kreuzberg: An open-source library developed by myself, optimized for speed and scalability.
• Docling: IBM’s machine learning-powered solution, known for advanced understanding but resource-heavy.
• MarkItDown: Microsoft’s Markdown converter, suitable for simple document formats.
• Unstructured: An enterprise-grade processing framework, designed for complex document workflows.

Testing Methodology

For a comprehensive overview, I utilized a dataset featuring 94 real-world documents—including PDFs, Word files, HTML pages, images, and spreadsheets—spanning five size categories from tiny files (<100KB) to massive documents (>50MB). The documents covered six languages, such as English, Hebrew, Chinese, and Japanese, to assess multilingual processing.

Benchmarking was conducted in a controlled environment using CPU-only processing to ensure fair comparisons. Metrics gathered included processing speed (files per second), memory consumption, success rates, and installation footprint.

Key Findings

Performance and Speed

• Kreuzberg emerged as the standout, processing over 35 files per second across all document types with consistent reliability.
• Unstructured demonstrated solid performance, particularly excelling in handling complex layouts, though at a moderate speed.
• MarkItDown performed well on straightforward documents like Markdown and simple PDFs but faltered on larger, more intricate files.
• Docling, despite its advanced capabilities, often took upwards of 60 minutes per file, with frequent timeouts on medium-sized documents, limiting