Multi-column pages that read in the wrong order

Sometimes you have data that flows over multiple columns, or pages, or just isn’t arranged in a “normal” top-to-bottom way. This page has three columns of text with two tables buried in them — and a plain extract_text() walks straight across all three columns at once, interleaving sentences that have nothing to do with each other.

from natural_pdf import PDF

pdf = PDF("pdfs/multicolumn.pdf")
page = pdf.pages[0]
page.show()

Look at the first few hundred characters of a naive extraction — each line jumps from column to column:

print(page.extract_text(layout=True)[:500])

This is some text             25     8115      XXX21        6783

          This is some text             26     8448      XXX22        7116
                                        27     8781      XXX23        7449
          This is some text             28     9114      XXX24        7782
                                        29     9447      XXX25        8115

The fix is to reflow the page: cut out each column as a region, then paste the pieces back together end-to-end so “below” means “further along in reading order.”

Cut the page into columns

page.region() cuts a rectangle by coordinates. Three vertical slices, each a third of the page wide:

left = page.region(left=0, right=page.width/3, top=0, bottom=page.height)
mid = page.region(left=page.width/3, right=page.width/3*2, top=0, bottom=page.height)
right = page.region(left=page.width/3*2, right=page.width, top=0, bottom=page.height)
page.highlight(left, mid, right)

Stack them with a Flow

A Flow connects separate regions (or whole pages) vertically or horizontally. Stacking the three columns vertically turns the page into one long single-column document:

from natural_pdf.flows import Flow

stacked = [left, mid, right]
flow = Flow(segments=stacked, arrangement="vertical")
flow.show()

Now any spatial operation — “find something below this” — follows reading order, even when that means jumping from the bottom of one column to the top of the next:

region = (
    flow
    .find('text:contains("Table one")')
    .below(
        until='text:contains("Table two")',
        include_endpoint=False
    )
)
region.show()

Pull out every table

The tables sit under bold headers. Find the headers first — the width > 10 filter skips some tiny empty boxes that also register as bold:

flow.find_all('text[width>10]:bold').show()

Then take each header and sweep down until the next bold header or the closing paragraph, whichever comes first. The | in the selector means “either of these”:

regions = (
    flow
    .find_all('text[width>10]:bold')
    .below(
        until='text[width>10]:bold|text:contains("Here is a bit")',
        include_endpoint=False
    )
)
regions.show()

Each region is one table, so extract_table() on a region can’t grab the wrong one:

regions[0].extract_table().to_df()

	index	number
0	1	123
1	2	456
2	3	789
3	4	1122
4	5	1455
5	6	1788
6	7	2121
7	8	2454
8	9	2787
9	10	3120
10	11	3453
11	12	3786
12	13	4119
13	14	4452
14	15	4785
15	16	5118
16	17	5451
17	18	5784
18	19	6117
19	20	6450
20	21	6783
21	22	7116
22	23	7449
23	24	7782
24	25	8115
25	26	8448
26	27	8781
27	28	9114
28	29	9447
29	30	9780
30	31	10113
31	32	10446
32	33	10779
33	34	11112
34	35	11445
35	36	11778
36	37	12111
37	38	12444
38	39	12777

And if the tables share a structure, combine them into one DataFrame:

import pandas as pd

dfs = regions.apply(lambda region: region.extract_table().to_df())
merged = pd.concat(dfs, ignore_index=True)
merged

	index	number
0	1	123
1	2	456
2	3	789
3	4	1122
4	5	1455
...	...	...
73	XXX35	11445
74	XXX36	11778
75	XXX37	12111
76	XXX38	12444
77	XXX39	12777

78 rows × 2 columns

If your columns aren’t this tidy

This page splits into exact thirds, so hardcoded coordinates work. When column widths vary — or you can’t hand-measure every document in a stack — page.analyze_layout() can detect text blocks for you (it downloads YOLO weights on first use), and the detected regions feed into a Flow the same way. The same Flow trick also spans pages: pass pdf.pages as the segments and a table that breaks across a page boundary becomes one table.