chrisryugj/kordoc

What it solves

Kordoc is a comprehensive document processing toolkit designed to handle the "document hell" of Korean government and administrative offices. It solves the problem of extracting, analyzing, and generating documents in formats that are traditionally difficult to parse, specifically HWP (Hangul Word Processor) versions 3.x and 5.x, HWPX, HWPML, PDF, XLS, XLSX, DOCX, and images (PNG/JPG/WebP).

How it works

Kordoc converts these various formats into Markdown, which is the ideal format for LLMs to read and analyze. It uses a combination of specialized parsers, a local CPU-based OCR engine (PP-OCRv5 korean) for scanned documents and images, and a custom reflow engine to render documents back into HWPX without needing the Hancom Office software installed on the server. It also supports a "round-trip" workflow where Markdown can be edited and patched back into the original binary files while preserving the original formatting.

Who it’s for

• AI Developers: Those building RAG (Retrieval-Augmented Generation) systems or AI agents that need to process Korean administrative documents.
• Government Employees/Administrators: People dealing with high volumes of official documents and reports.
• AI Agent Users: Users of Claude Desktop, Cursor, and other MCP-compatible clients who want to integrate document parsing and generation directly into their AI workflow.

Highlights

• Multi-format Parsing: Supports HWP, HWPX, PDF, DOCX, and Excel files, converting them to Markdown with high fidelity.
• Complex Table Recovery: Specialized in restoring merged and nested tables, including those in PDFs without visible borders.
• Government Standard Generation: An engine capable of generating official Korean government documents (reports, drafts, notices) following strict formatting rules.
• Local OCR: Built-in OCR for scanned PDFs and images that runs locally on the CPU without requiring API keys.
• RAG-Ready Output: Provides structured chunks with breadcrumb hierarchies and actual page numbers for precise citations.
• MCP Integration: Seamlessly connects with AI agents (Claude, Cursor, etc.) via the Model Context Protocol (MCP) to read and create documents.
• Privacy Masking: Automatic detection and masking of personal information (resident registration numbers, phone numbers, etc.) while preserving document formatting.