On-premise document parser
Turn HWP documents into RAG-ready data,
on your own servers.
Luminir Document Parser is an on-premise document parser built by LEMONG. It finds titles, text, tables and figures in HWP/HWPX, PDF, Office and scanned documents, and outputs them in reading order as HTML and JSON. The recognition models run on your server alongside the parser.
How it works
Accept files
12 file types: HWP/HWPX, PDF, Word, PowerPoint, Excel and images. No conversion needed first.
Layout analysis
Detects titles, body text, tables, figures and page headers on every page, and decides the reading order.
Content recognition
Reads text as text, tables as rows and columns (merged cells included), and formulas in PDFs and scans as LaTeX.
Output
HTML, plus JSON with the type, page number, reading order and coordinates of every element (sheet and cell range for Excel). Headers and page numbers are labeled, so they are easy to drop.
Also included: Native HWP/HWPX parsing, Heading detection, Header and page number labels, Source coordinates for every element, Resumable batch conversion, Closed-network deployment.
Public benchmarks (self-measured)
| Benchmark | Data | Score | Comparison | Date |
|---|---|---|---|---|
| opendataloader-bench | Published by Hancom · 200 English pages | 0.914 | Higher than the public #1 score (0.907) | 2026-10-12 |
| KoOCR-Bench | schift-io v1.2 · 604 real-world Korean pages | 0.759 | Highest among installable engines on the public leaderboard | 2026-10-12 |
| DP-Bench | Published by Upstage · 200 English pages · table TEDS | 0.962 | Higher than the public #1 score (0.961) | 2026-09-23 |
| OmniDocBench v1.5 | 1,651 pages · table TEDS | 0.934 | — | 2026-10-23 |
All scores are self-measured by LEMONG using each benchmark’s public data and scoring code. Measurement dates are shown.
Input
- HWP/HWPX
- Korean word processor files, parsed natively
- Digital and scanned
- Word, PowerPoint
- DOC, DOCX, PPT, PPTX
- Excel
- XLS and XLSX, as tables
- Images
- PNG, JPG, JPEG
Output
- HTML
- Separates headings, tables and figures; opens in any browser
- JSON
- Structured data with the type and position of every element
- REST API
- Upload and retrieve results programmatically
Your documents stay on your servers
- On-premise and closed networksInstalled as containers on a Linux server with an NVIDIA GPU.
- Runs without external AI servicesLayout and recognition models run on your server, so it works inside your internal network without external AI services.
- Access control and auditAdmin and user roles, encrypted storage of original documents, and audit logs of document uploads, views and downloads.
Talk to us
Tell us what documents you handle, how many, and your server setup (closed network, GPU). We reply in English or Korean.
support@lemong.ai · 02-6958-6868 · LEMONG Co., Ltd.