DEEP Parser
Documents, rebuilt as data your AI can trust.
Keeps the structure of a document intact, including tables, charts, merged cells, and multi-column layouts, and turns it into data that search and agents can use right away.
When parsing breaks, everything after it breaks.
Answers from RAG and agents often wobble because of the input document more than the model. That's why KDL doesn't stop at parsing. It builds the entire process with you, from the moment a document arrives to the moment an answer goes out.
Document
Takes PDF, HWP, Office documents, and scans as they are, with no conversion.
01Document
Takes PDF, HWP, Office documents, and scans as they are, with no conversion.
02Parser
Reads multi-column layouts, table headers, and captions in their original structure. If this step breaks, every step after it breaks too.
03Chunking
Splits by clause, paragraph, and table rather than by character count, so meaning is not cut in half.
04Embedding
Attaches the title, page, date, and document type to each chunk and converts it into a vector.
05VectorDB
Keeps the index inside the customer's server. It runs the same way on an air-gapped network.
06Retrieval
Finds documents that answer the current question, not merely similar ones, using their metadata too.
07Re-ranking
Reorders the retrieved evidence by how well it fits the question.
08Generation
Attaches the original location of the source for each answer, so you can check it right away.
Reads a page the way people do.
Splits titles, paragraphs, tables, pictures, charts, and formulas into elements and sets the reading order. Even when each document is laid out differently, it organizes them into the same structure using 23 layout tags.
23
Layout tags
- Title
- Section header
- Text
- List item
- Caption
- Footnote
- Table
- Picture
- Chart
- Formula
- Flowchart
01 / 08NASA Facts
Tables
Tables inside tables, merged cells intact.
Keeps the row and column relationships intact, including tables nested inside tables and merged cells. No value is dropped or overlapped, so evaluation sheets and administrative forms become data with their original structure.
Charts and drawings
Charts read as numbers, not pictures.
Instead of leaving charts, graphs, and drawings as pictures, it moves the axes, legends, and relationships between values into data. Information that existed only as an image can now be searched and analyzed.
Chunks and metadata
Chunks that keep their meaning.
Splits by clause, paragraph, and table rather than by character count. It keeps metadata such as title, page, date, and document type, so it finds the document that fits the current question rather than a similar one.
Output your systems already read.
Takes PDF, HWP, Office documents, and scans and returns them as HTML, JSON, or Markdown. The results feed straight into ERP, EDMS, RPA, and RAG and agent pipelines.
In
- HWP
- Word
- Excel
- PowerPoint
- Scans
- Images
Out
- HTML
- JSON
- Markdown
Into
- ERP
- EDMS
- RPA
- RAG
- LLM
- AI Agent
#1 of 33 on ParseBench.
LlamaIndex’s ParseBench measures document structuring ability with about 2,000 pages of real corporate documents from insurance, finance, and the public sector. On the official Hugging Face leaderboard, KDL’s 1.2B parser model ranked first among 33 models.
Overall score, official leaderboard
1.2B
Parameters
Designed to run on a single GPU, so it can operate inside an air-gapped network.
1/30
Size of the previous open-source leader
About one-thirtieth the size of the previous open-source leader (35.1B).
Open
Weights
The weights are public, so customers can check and review the performance themselves.
Case · Gyeonggi Provincial Government
A province-wide AI platform that started with documents.
Gyeonggi Provincial Government pursued a generative AI platform project in 2025. It was the first case among provincial-level local governments, and more than 90% of its internal documents were HWP or PDF. So it solved document structuring before the model, and DEEP Parser was deployed there.
KRW 13.1B
Generative AI platform program, 2025
90%+
Internal documents in HWP or PDF
97.3%
Item recognition accuracy, internal test
Security
Runs inside your closed network.
Runs just as well in on-premise and closed-network environments fully cut off from external networks. Documents and models never leave your organization.
GS certified
Software quality certification (GS) Grade 1
ISO 9001
Quality management certification
Women-owned business
Women-Owned Business Certification
K-Innovative Product
Innovative Procurement Product by PPS
Choose how it runs
On-Premise
Installed directly on your own servers and operated there.
Fits when
- Servers are ready
- External internet cannot be used
- Security matters most
Private Cloud
Operated in a dedicated cloud with data kept separate.
Fits when
- You want a closed-network cloud
- A dedicated environment and data separation are required
- You want to hand off server operations
DEEP Hub
Install dedicated software as a package, even without an internal server.
Fits when
- You have no internal servers
- Buying servers all at once is a burden
- You want to start using the software right away
SaaS
Start right away in the cloud, with no installation.
Fits when
- You want to use it right away with no installation
- You can use the internet
- You need a reasonable price
Bring us your hardest document.
Share the documents where tables break and the order gets scrambled. Send an inquiry and an engineer in charge will review it and contact you directly.