Document Question Answering¶

Natural PDF includes document QA functionality that allows you to ask natural language questions about your PDFs and get relevant answers. This feature uses LayoutLM models to understand both the text content and the visual layout of your documents.

Setup¶

Let's start by loading a sample PDF to experiment with question answering.

In [1]:

from natural_pdf import PDF

# Path to sample PDF
pdf = PDF("https://github.com/jsoma/natural-pdf/raw/refs/heads/main/pdfs/0500000US42001.pdf")

# Display the first page 
page = pdf.pages[0]
page.show()

from natural_pdf import PDF # Path to sample PDF pdf = PDF("https://github.com/jsoma/natural-pdf/raw/refs/heads/main/pdfs/0500000US42001.pdf") # Display the first page page = pdf.pages[0] page.show()

Out[1]:

Basic Usage¶

Here's how to ask questions to a PDF page:

In [2]:

# Ask a question about the entire document
page.ask("How many votes did Harris and Waltz get?")

# Ask a question about the entire document page.ask("How many votes did Harris and Waltz get?")

Device set to use mps:0

Out[2]:

{'answer': '148',
 'confidence': 0.9995507001876831,
 'start': 20,
 'end': 20,
 'found': True,
 'page_num': 0,
 'source_elements': <ElementCollection[TextElement](count=1)>}

In [3]:

page.ask("Who got the most votes for Attorney General?")

page.ask("Who got the most votes for Attorney General?")

Out[3]:

{'answer': 'DEM EUGENE DEPASQUALE',
 'confidence': 0.9180722236633301,
 'start': 63,
 'end': 63,
 'found': True,
 'page_num': 0,
 'source_elements': <ElementCollection[TextElement](count=1)>}

In [4]:

page.ask("Who was the Republican candidate for Attorney General?")

page.ask("Who was the Republican candidate for Attorney General?")

Out[4]:

{'answer': 'LIB ROBERT COWBURN',
 'confidence': 0.21592436730861664,
 'start': 67,
 'end': 67,
 'found': True,
 'page_num': 0,
 'source_elements': <ElementCollection[TextElement](count=1)>}

Asking questions to part of a page questions¶

You can also ask questions to a specific region of a page*:

In [5]:

# Get a specific page
region = page.find('text:contains("Attorney General")').below()
region.show()

# Get a specific page region = page.find('text:contains("Attorney General")').below() region.show()

Out[5]:

In [6]:

region.ask("How many write-in votes were cast?")

region.ask("How many write-in votes were cast?")

Out[6]:

{'answer': '498',
 'confidence': 0.9988918304443359,
 'start': 17,
 'end': 17,
 'found': True,
 'region': <Region bbox=(0, 553.663, 612, 792)>,
 'page_num': 0,
 'source_elements': <ElementCollection[TextElement](count=1)>}

Asking multiple questions¶

In [7]:

import pandas as pd

questions = [
    "How many votes did Harris and Walz get?",
    "How many votes did Trump get?",
    "How many votes did Natural PDF get?",
    "What was the date of this form?"
]

# You can actually do this but with multiple questions
# in the model itself buuuut Natural PDF can'd do it yet
results = [page.ask(q) for q in questions]

df = pd.json_normalize(results)
df.insert(0, 'question', questions)
df

import pandas as pd questions = [ "How many votes did Harris and Walz get?", "How many votes did Trump get?", "How many votes did Natural PDF get?", "What was the date of this form?" ] # You can actually do this but with multiple questions # in the model itself buuuut Natural PDF can'd do it yet results = [page.ask(q) for q in questions] df = pd.json_normalize(results) df.insert(0, 'question', questions) df

Out[7]:

	question	answer	confidence	start	end	found	page_num	source_elements
0	How many votes did Harris and Walz get?	148	0.999671	20	20	True	0	[<TextElement text='148' font='Helvetica' size...
1	How many votes did Trump get?	348	0.310203	22	22	True	0	[<TextElement text='348' font='Helvetica' size...
2	How many votes did Natural PDF get?	November 5, 2024	0.237136	3	3	True	0	[<TextElement text='November 5...' font='Helve...
3	What was the date of this form?	November 5, 2024	0.792696	3	3	True	0	[<TextElement text='November 5...' font='Helve...

Next Steps¶

Now that you've learned about document QA, explore:

• Element Selection: Find specific elements to focus your questions.
• Layout Analysis: Automatically detect document structure.
• Working with Regions: Define custom areas for targeted questioning.
• Text Extraction: Extract and preprocess text before QA.