[Python] How can I accurately extract Korean text from image-based PDFs for use with LLMs?

I’m a university student currently leading an AI (LLM) project at my school.

I’m working with PDF files that are mostly scanned images containing Korean text. My goal is to extract this text and store it in a vector database for use with a large language model (LLM). However, I’m facing significant challenges with OCR and text extraction.

I’ve tried various Python libraries—such as pymupdf, Llamaparse, pdf2image, pdfplumber, and pdfminer—but most of them either fail to read image-based PDFs or do not correctly recognize Korean characters. This has become a major roadblock in integrating OCR and LLM training into our service workflow.

What I tried: I used the tools mentioned above expecting accurate text extraction from Korean image-based PDFs, which would allow for vectorization and LLM use. Unfortunately, the results were either unreadable or missing Korean character support entirely.

If anyone has recommendations for tools, workflows, or models that support high-quality Korean OCR from scanned PDFs, I would be extremely grateful for your advice.

Thank you, and wishing you a great day!

Continue reading...