In [ ]:

# Install required packages
!pip install --upgrade --quiet 'natural-pdf[ai]>=0.5.0'

print('✓ Packages installed!')

Slides: slides.pdf

Multi-page flows¶

Sometimes you have data that flows over multiple columns, or pages, or just... isn't arranged in a "normal" top-to-bottom way.

In [1]:

from natural_pdf import PDF

pdf = PDF("https://github.com/jsoma/natural-pdf/raw/refs/heads/main/pdfs/multicolumn.pdf")
page = pdf.pages[0]
page.show(width=800)

Out[1]:

No description has been provided for this image

Natural PDF deals with these through reflowing pages, where you grab specific regions of a page and then paste them back together either vertically or horizontally.

In this example we're splitting the page into three columns.

In [2]:

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)
mid.show()

Out[2]:

Now let's stack them on top of each other.

In [3]:

from natural_pdf.flows import Flow

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

Out[3]:

In [4]:

flow.show(in_context=False)

Out[4]:

Now any time we want to use spatial comparisons, like "find something below this," it just works.

In [5]:

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

Out[5]:

It works for text, it works for tables, it works for anything. Let's see how we can get both tables on the page.

First we find the bold headers – we need to say width > 10 because otherwise it pulls some weird tiny empty boxes.

In [6]:

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

Out[6]:

Then we take each of those headers, and go down down down until we either hit another bold header or the "Here is a bit more text" text.

In [7]:

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()

Out[7]:

Now we can use .extract_table() on each individual region to give us a bunch of tables.

In [8]:

regions[1].extract_table().to_df()

Out[8]:

	index	number
0	XXX1	123
1	XXX2	456
2	XXX3	789
3	XXX4	1122
4	XXX5	1455
5	XXX6	1788
6	XXX7	2121
7	XXX8	2454
8	XXX9	2787
9	XXX10	3120
10	XXX11	3453
11	XXX12	3786
12	XXX13	4119
13	XXX14	4452
14	XXX15	4785
15	XXX16	5118
16	XXX17	5451
17	XXX18	5784
18	XXX19	6117
19	XXX20	6450
20	XXX22	7116
21	XXX23	7449
22	XXX24	7782
23	XXX25	8115
24	XXX26	8448
25	XXX27	8781
26	XXX28	9114
27	XXX29	9447
28	XXX30	9780
29	XXX31	10113
30	XXX32	10446
31	XXX33	10779
32	XXX34	11112
33	XXX35	11445
34	XXX36	11778
35	XXX37	12111
36	XXX38	12444
37	XXX39	12777

Layout analysis and magic table extraction¶

Similar to how we have feelings about what things are on a page - headers, tables, graphics – computers also have opinions! Just like some AI models have been trained to do things like identify pictures of cats and dogs or spell check, others are capable of layout analysis - YOLO, surya, etc etc etc. There are a million! TATR is one of the useful ones for us, it's just for table detection.

But honestly: they're mostly trained on academic papers, so they aren't very good at the kinds of awful documents that journalists have to deal with. And with Natural PDF, you're probably selecting text[size>12]:bold in order to find headlines, anyway. But if your page has no readable text, they might be able to provide some useful information.

Let's start with YOLO, the default.

In [9]:

from natural_pdf import PDF

pdf = PDF("https://github.com/jsoma/natural-pdf/raw/refs/heads/main/pdfs/needs-ocr.pdf")
page = pdf.pages[0]

In [10]:

page.analyze_layout('yolo')
(
    page
    .find_all('region')
    .show(group_by='type', width=800)
)

image 1/1 /var/folders/25/h3prywj14qb0mlkl2s8bxq5m0000gn/T/tmp_ku_yary/temp_layout_image.png: 1024x800 2 titles, 3 plain texts, 2 abandons, 1 table, 1445.2ms

Speed: 5.8ms preprocess, 1445.2ms inference, 1.4ms postprocess per image at shape (1, 3, 1024, 800)

Out[10]:

In [11]:

page.find('table').apply_ocr()
text = page.extract_text()
print(text)

Using CPU. Note: This module is much faster with a GPU.

/Users/soma/Development/natural-pdf/.venv/lib/python3.11/site-packages/torch/utils/data/dataloader.py:775: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.
  super().__init__(loader)

Statute Description Level Repeat?
4.12.7 Unsanitary Working Conditions Critical
5.8.3 Inadequate Protective Equipment: Serious
6.3.9 Ineffective Injury Prevention: Serious
7.1.5 Failure to Properly Store Hazardous Materials: Critical
8.9.2 Lack of Adequate Fire Safety Measures. Serious
9.6.4 Inadequate Ventilation Systems. Serious
10.2.7 Insufficient Employee Training for Safe Work Practices. Serious

Better layout analysis with tables¶

Let's see what TATR - Microsoft's table transformer – finds for us.

In [12]:

page.analyze_layout('tatr')
page.find_all('region').show(group_by='type', width=800)

Out[12]:

There's just so much stuff that TATR is finding that it's all overlapping.

For example, we can just look at one piece at a time.

In [13]:

# table-cell
# table-row
# table-column
page.find_all('region[type=table-row]').expand(-2).show(crop=True, width=800)

Out[13]:

In [14]:

# Grab all of the columns
cols = page.find_all('region[type=table-column]')

# Take one of the columns and apply OCR to it
cols[2].apply_ocr()
text = cols[2].extract_text()
print(text)

Level
Critical
Serious
Serious
Critical
Serious
Serious
Serious

/Users/soma/Development/natural-pdf/.venv/lib/python3.11/site-packages/torch/utils/data/dataloader.py:775: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.
  super().__init__(loader)

In [15]:

len(cols[2].find_all('text[source=ocr]'))

Out[15]:

In [16]:

page.find('table').show()

Out[16]:

In [17]:

data = page.find('table').extract_table()
data

Out[17]:

TableResult(rows=9…)

YOLO¶

In [18]:

page.analyze_layout()
page.find_all('region').show(group_by="type")

image 1/1 /var/folders/25/h3prywj14qb0mlkl2s8bxq5m0000gn/T/tmpm2fdyzg6/temp_layout_image.png: 1024x800 2 titles, 3 plain texts, 2 abandons, 1 table, 775.6ms

Speed: 4.5ms preprocess, 775.6ms inference, 1.2ms postprocess per image at shape (1, 3, 1024, 800)

Out[18]:

In [19]:

page.find("region[type=table]").apply_ocr()

Out[19]:

<Region type='table' source='detected' bbox=(99.85887908935547, 815.1997680664062, 1146.6968994140625, 1153.8369140625)>

In [20]:

text = page.extract_text()
print(text)

Statute Description Level Repeat?
4.12.7 Unsanitary Working Conditions Critical
5.8.3 Inadequate Protective Equipment: Serious
6.3.9 Ineffective Injury Prevention: Serious
7.1.5 Failure to Properly Store Hazardous Materials: Critical
8.9.2 Lack of Adequate Fire Safety Measures. Serious
9.6.4 Inadequate Ventilation Systems. Serious
10.2.7 Insufficient Employee Training for Safe Work Practices. Serious

In [21]:

from natural_pdf.analyzers.guides import Guides

table_area = page.find("region[type=table]")
guides = Guides(table_area)
guides.horizontal.from_lines()
guides.vertical.from_content(["Description", "Level", "Repeat"])
guides.vertical.snap_to_whitespace()
guides.show()

Out[21]:

In [22]:

guides.extract_table().to_df()

Out[22]:

	Statute	Description	Level	Repeat?
0	4.12.7	Unsanitary Working Conditions	Critical	None
1	5.8.3	Inadequate Protective Equipment:	Serious	None
2	6.3.9	Ineffective Injury Prevention:	Serious	None
3	7.1.5	Failure to Properly Store Hazardous Materials:	Critical	None
4	8.9.2	Lack of Adequate Fire Safety Measures.	Serious	None
5	9.6.4	Inadequate Ventilation Systems.	Serious	None
6	10.2.7	Insufficient Employee Training for Safe Work P...	Serious	None