Scientific PDF Parsing for RAG Pipelines Fails Silently on Equations and Tables

Developers building retrieval-augmented generation (RAG) systems over scientific papers face a persistent problem: standard PDF parsers struggle with two-column layouts, embedded equations, and nested tables common in academic literature. Unlike business documents, scientific PDFs often encode math as images or vector graphics, forcing OCR tools to reconstruct complex notation rather than simply read text. A developer tested three extraction pipelines — MinerU, Docling, and Marker — across 500 papers containing 120 formulas, finding that MinerU achieved roughly 80% accuracy on formula extraction while Docling faltered on uncommon notation. The critical concern is not the failure rate but the failure mode: errors appear silently, producing plausible-looking but incorrect values that can enter a RAG pipeline as untagged hallucinated facts. The developer concluded that even the best-performing parser requires a dedicated verification layer to flag low-confidence extractions before they reach an LLM-powered application.