Document understanding · specification and research foundation
Reading a page as a whole.
Keeping extracted headings, paragraphs, figures, and tables attached to their source regions, reading order, and review state.
01
A document page is more than a flat string of words.
A flat text export can mix two columns, separate a caption from its figure, or lose the row and column around a table value. A reader then has no clear way to check where an extracted statement came from.
aKriti is a document vision-language foundation being designed around grounded structure. Its draft representation keeps page regions, block roles, reading order and source references together so future tools can make extraction inspectable.
02
Select a page region or its outline row.
Explore the example page and its outline together. Selecting a block connects its visible region to a plain-language explanation and the fields recorded for it.
This is an authored educational example; no live parse or model inference runs here. The regions, text and illustrative uncertainty are supplied for the lesson, not measured output from aKriti.
Read a page · authored example
A page and its structure
Choose a region to connect the visible page, its text and its position. These two pages are human-authored teaching examples, not aKriti model output.
This teaching example shows how source regions can stay linked to their fields. It does not run a parser, save document changes or report model accuracy.
Page 1 · Title selected · The page title
Source page
1 / 2Choose a marked region, or use the reading-order links.
Page 1 · A note and a sketch. Text and region boxes use the same authored source coordinates.
Title · source region
The page title
A title is a region with its own identity and location. Its position and role matter alongside the words.
Visible source text
A walk, kept on paper
See the partial field record
Selected fields follow the draft aKritiDoc coordinate and reading-order names. This authored excerpt is not a complete or validated aKritiDoc.
{
"example": "Partial draft-v0 field excerpt from a human-authored website specimen. This is not a complete or validated aKritiDoc.",
"page": {
"page_id": "page_notebook_01",
"page_index": 0,
"width": 1000,
"height": 1400,
"unit": "px",
"reading_order": [
"blk_notebook_title",
"blk_notebook_note",
"blk_notebook_figure",
"blk_notebook_stamp"
]
},
"block": {
"block_id": "blk_notebook_title",
"type": "title",
"bbox": {
"x": 0.1,
"y": 0.08,
"w": 0.8,
"h": 0.1,
"space": "normalized",
"page_id": "page_notebook_01"
},
"spans": [
{
"span_id": "span_notebook_title",
"text": "A walk, kept on paper",
"language": "en",
"script": "Latin",
"bbox": {
"x": 0.1,
"y": 0.105,
"w": 0.8,
"h": 0.045,
"space": "normalized",
"page_id": "page_notebook_01"
}
}
],
"metadata": {
"example_provenance": "human-authored website specimen"
}
}
}03
What stays attached to every block.
Every region needs a name, a role and a place on the page. The written note, sketch and small COPY mark are different kinds of content, even when each contains words.
Select one in the example. Its outline entry, source text and field record refer to the same region. That connection is what lets a reader check an extracted statement against the original.
04
Columns, figures, and tabular cells.
The order of blocks on the page is a separate record from their geometric positions. A two-column page should not be read by blindly crossing the gutter at every line.
In the draft, a table records rows, columns and individual cells. Cells have row and column positions, spans, text and their own boxes. This preserves the surrounding structure needed to interpret a value and check its source.
05
Human review status is not the same as model confidence.
A system saying it is confident is different from a person checking the page. A useful review view keeps both the original and the reason for a check within reach.
On page two, select the inventory. Its review note asks you to compare the quantity 18 with the source. The example leaves that number unchanged; it does not invent a confidence score or apply a correction.
06
Schema specification & artifact flow.
aKritiDoc is the planned shared format for a document, its pages, regions and source links. Parsing tools and review tools should be able to use that same structure instead of passing around disconnected text.
The example’s optional field record shows a small part of that idea. It follows one draft convention for positions and reading order; it is not a complete aKritiDoc. The repository drafts still need reconciliation and an executable validator.
07
Current status and evidence gates.
The repository provides specification and architecture groundwork. A working parser, executable schema, review workflow and reproducible evaluation still need their own implementation and evidence.
The next useful proof is a real document passing through parsing, source checks and review. Accuracy, speed and language support will need reproducible tests. This page presents the design direction and an authored lesson, not those results.