Why does parsing scientific papers for RAG still break on equations and tables?

Short summary

Scientific PDF parsing for RAG pipelines fails silently on equations, nested tables, and two-column layouts—producing plausible-looking but incorrect output that becomes hallucinated facts downstream. Testing MinerU, Docling, and Maker across 500 papers showed MinerU at ~80% formula extraction accuracy, but a verification pass still caught 80 mismatches. The key gap is section-level confidence scoring: knowing which parts of a document to trust and which to double-check before feeding into an LLM pipeline.

• •Silent extraction errors in scientific PDFs are more dangerous than loud failures
• •MinerU leads raw extraction at ~80% formula accuracy but still needs verification
• •Section-level confidence scoring is the missing piece for safe RAG over papers

Generated with AI, which can make mistakes.