Comprehensive Evaluation of Python Text Extraction Libraries in 2025: An In-Depth Benchmarking Report

In the rapidly evolving landscape of document processing, selecting the optimal Python library for text extraction can be a daunting task. To aid developers and data scientists, we present an impartial and thorough comparison of four prominent Python-based text extraction tools, based on rigorous testing conducted in 2025. This analysis scrutinizes their performance, reliability, and suitability across a diverse array of real-world documents.

Overview of the Benchmark

Our evaluation examines four leading libraries:

• Kreuzberg – an open-source, lightweight solution designed for high-speed processing
• Docling – IBM’s machine learning-powered extraction framework
• MarkItDown – Microsoft’s Markdown-focused converter for simplified documents
• Unstructured – a versatile platform aimed at enterprise-level document handling

The testing protocol encompassed an extensive dataset of 94 documents, including PDFs, Word files, HTML pages, images, and spreadsheets. These files ranged from small (<100KB) to very large (>50MB), encompassing multiple languages such as English, Hebrew, German, Chinese, Japanese, and Korean. To ensure fairness, all operations were conducted on CPU-only systems, with meticulous tracking of processing times, memory consumption, and success rates.

Key Findings from the 2025 Results

Performance and Speed

Our tests revealed striking disparities among the libraries:

• Kreuzberg emerged as the fastest, processing over 35 documents per second across the board, excelling especially with large or complex files.
• Unstructured provided robust reliability but at moderate speeds, making it suitable for scenarios where consistency is paramount.
• MarkItDown handled straightforward, simple documents efficiently but struggled with complex or sizable files, particularly those exceeding 10MB.
• Docling, despite its advanced ML capabilities, often required more than 60 minutes per document, with frequent timeouts on medium-sized files.

Installation Footprint and Resource Usage

The size of the library installation directly impacts deployment options:

• Kreuzberg boasts a compact size of approximately 71MB with only 20 dependencies, making it ideal for resource-constrained environments.
• Unstructured is larger, around 146MB, with 54 dependencies, suitable for enterprise contexts.
• MarkItDown has a moderate footprint (~251MB) but includes heavy dependencies like ONNX.
• **Docling