Every document,
trusted by AI.

Luminir Document Parser is a document parser built by LEMONG that supports both on-premise and cloud deployment. It finds titles, body text, tables and figures in HWP, HWPX, PDF, Word, PowerPoint, Excel and scanned image documents, and converts them into HTML and JSON in reading order. It is used to build data for RAG (AI that searches documents to answer) systems.

It is used for Q&A on regulations and guidelines, batch conversion of accumulated HWP documents, turning statistical tables and reports into data, and showing where an AI answer came from in the original. You can feed rulebooks and manuals into a search system with their clauses and tables intact, or move statistical tables full of merged cells into Excel or a database with their rows and columns as they are.

You can check its performance in numbers. Luminir Document Parser publishes its scores on 4 public benchmarks, including the Korean public-document benchmark KoOCR-Bench, together with the measurement conditions, and we will show you the results on your own documents.

It accepts 12 file types: HWP and HWPX, PDFs with text and scanned PDFs, Word (DOC, DOCX), presentations (PPT, PPTX), Excel (XLS, XLSX) and images (PNG, JPG, JPEG), as they are, with no separate conversion for each format.

Yes, just put in HWP and HWPX files as they are. It reads the paragraphs and tables of Hangul files directly. Batch conversion of many files at once is supported, and if it stops partway, it resumes from where it stopped.

Yes. For scanned PDFs and photos, it finds table regions in the page image and reads them as rows and columns. On KoOCR-Bench, made of 604 pages of real Korean documents such as the Official Gazette and statistical tables, we measured an overall score of 0.759 (self-measured). We will show you the results on your own documents.

Yes. Tables with merged cells are exported as HTML tables that carry the row and column merge information. On the public benchmark DP-Bench, we measured a table structure and content score (TEDS) of 0.962 (self-measured).

In HTML and JSON. The HTML is a document that separates headings, tables and figures and opens right in a browser, and the JSON holds the type, page number, reading order and coordinates of each element (sheet and cell range for Excel). A REST API covers everything from upload to retrieving results.

They are kept and labeled by type. Headers, footers and page numbers are marked separately, so they are easy to filter out if you want them out of your search data.

Formula regions in PDFs and scanned documents are read as LaTeX (a notation for formulas).

Yes. Every element has a page number and coordinates, so the region can be highlighted on the original document (sheet and cell range for Excel). The element position score on KoOCR-Bench is 0.976 (self-measured).

They are self-measured by LEMONG using the data and scoring code each benchmark has published, as they are. The measurement conditions are published on the benchmark details page.

On the Korean public-document benchmark KoOCR-Bench (604 pages, including the Official Gazette, statistical tables and reports), we measured an overall score of 0.759 (self-measured). It is the highest score among the installable engines on the public leaderboard.

Yes. It first finds the regions on the whole page, decides the reading order, and then reads each region. On opendataloader-bench, made of 200 English pages, we measured a reading order score of 0.944 (self-measured).

Both are supported. On-premise, it is installed on your servers and works even on closed networks with no internet connection. In the cloud, documents are processed on LEMONG servers, so you do not need to prepare a server. We will advise you based on your security policy and document volume.

A Linux server with an NVIDIA GPU and a container (Docker) runtime. The recognition models are installed on your server too, so it runs inside your internal network without external AI services. We will recommend detailed server specs based on your document volume and number of concurrent users.

On-premise, the layout analysis and recognition models are installed together on your servers, so documents are processed inside them. It runs inside your internal network without external AI services. In the cloud, documents are processed on LEMONG servers.

Yes. It separates admin and user roles, stores original documents encrypted, and keeps audit logs of document uploads, views, downloads and deletions.

Through a REST API. File upload, conversion progress and result retrieval are all available through the API, so it can plug into your search or RAG system.

Yes. Batch conversion of many files at once is supported, and if it stops, it resumes from where it stopped. We will measure processing time on your own documents.

Consultation, then a test on your documents, then contract, installation and integration. In the consultation we check your document types and volume and your server environment, and then explain the timeline and terms.