Fix PDF OCR Text with MinerU & Qwen2.5-VL in n8n

Fix garbled PDF OCR text automatically—page by page—using MinerU + Qwen2.5-VL in n8n

If your PDFs come back with messy OCR (misplaced characters, missing words, wrong line breaks), this n8n workflow helps clean and verify the text by cross-checking it against the actual page images using MinerU and Qwen2.5-VL (qwen2.5vl:7b) via Ollama.

What this workflow does

• Reads a local PDF from disk and converts it into per-page images.
• Sends the PDF/page data to a local MinerU parsing API at http://127.0.0.1:8000/file_parse to generate a result URL.
• Downloads the MinerU output, loads the generated Markdown file, and extracts the OCR text.
• Pairs each page image (base64) with its corresponding OCR text, then builds a prompt for the Ollama generate API to correct mismatches against the image.
• Loops through pages, calling Ollama with a short delay between requests.
• Post-processes the model output by removing Chinese characters, then aggregates all corrected page results into a single final output.

Use cases

• Repair OCR output for scanned documents (forms, reports, or statements) where text extraction is unreliable.
• Normalize page-by-page text for searchable PDF pipelines that require higher OCR accuracy before indexing.
• Improve extracted text quality for SaaS operators building document ingestion workflows in n8n.

Technical details

• MinerU: local HTTP endpoint http://127.0.0.1:8000/file_parse, expects output in an expected directory.
• Ollama: http://192.168.1.102:11434/api/generate using model qwen2.5vl:7b.
• n8n nodes/steps: set, code, wait, split out, aggregate, and sticky note.
• Setup required: update input/output filesystem paths and ensure both services are reachable.