Skip to content

Guide

Parse dense Korean slides the way ADE does, on open weights

What LandingAI's Agentic Document Extraction (DPT-3) is known to do, what the public evidence implies about how it does it, and a reproducible pipeline that reaches for the same output shape with Qwen-family and other open vision-language models from Hugging Face.

How DPT-3 works Korean slides pipeline Model candidates

What this wiki is for

People read a slide deck and understand it: the title frames the page, the bullets argue a point, the chart proves it, and the footnote qualifies it. A language model given the same slide as pixels sees a wall of small Korean text, a table without gridlines, and a chart whose legend is three words long. This wiki is the working notebook for closing that gap without a vendor API.

  • Understand the target

    LandingAI's Parse, Extract and Ground APIs, the DPT-3 Pro and Verity models, and the exact JSON shape they return. What LandingAI built

  • Reconstruct the pipeline

    An evidence-graded reconstruction of DPT-3: documented, strongly implied, unknown. Every claim cites its source. How DPT-3 works

  • Build the equivalent

    Render, layout, per-block reading, slide context, verification, and ADE-shaped output, with reference Python code. Korean slides pipeline

  • Pick the models

    Qwen3-VL, Qwen3.5, Qwen3.8 and the specialised 1B parsers that beat them on OmniDocBench, with code and published results. Model candidates

  • Retrieve over the result

    Text and page-image embeddings that actually work on Korean, with the KoVRE benchmark numbers. Embeddings and retrieval

  • Size the hardware

    A100 40/80 GB in 2 or 4 cards, TPU, and M1 Max, with memory budgets and serving stacks. Hardware and serving

How to read the evidence

This wiki does not have access to LandingAI's models or code. Three chips mark the strength of every claim about their system:

Documented stated in LandingAI's docs, OpenAPI spec, changelog or blog, with a link. Implied a reasonable inference from documented behaviour; the reasoning is shown. Unknown not publicly knowable; stated so you do not mistake a guess for a fact.

Model quality numbers for Korean are a second caveat: no public benchmark measures Korean slide parsing, so every model page ends with how to measure it yourself. See the evaluation recipe.