Comparison of Eight PDF Parsers Across 14 Document-Parsing Capabilities

A developer evaluated eight PDF parsers and OCR models, including Chandra, MinerU, and Granite-Docling, across 14 distinct capabilities. The benchmark revealed that Datalab's Chandra model achieved a perfect 14/14 accuracy score but suffered from a high latency of 91 seconds per page on an L4 GPU. PDF parsing is a critical bottleneck in preparing data for Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) pipelines. This benchmark highlights the stark trade-offs between processing speed and accuracy, especially when dealing with complex elements like handwriting, tables, and mathematical formulas. While Chandra excelled at handling merged-cell tables, LaTeX, and cursive handwriting, most other parsers struggled heavily with handwriting or hallucinated text over document stains. LightOnOCR-1B emerged as a faster alternative, processing at 7.9 seconds per page, though it suffered from hallucinations and incomplete page parsing.