PlasmaPDF Quick Start Guide
PlasmaPDF is a Python library for converting from txt spans to x-y positioned tokens in the PAWLs format. It is a utility library used in OpenContracts and PdfRedactor.
Installation
To install PlasmaPDF, use pip:
pip install plasmapdf
Understanding PdfDataLayer: A Bridge Between Span and PDF Coordinates
The PlasmaPDF's PdfDataLayer is a solution to a complex problem: maintaining perfect synchronization between plain text spans and their physical locations in PDFs. Let's break it down:
Core Concept / Use Case
PdfDataLayer is designed to keep span-based annotations consistent with the underlying PDF tokens such that's it's easy to convert between the two. In today's LLM-powered world, one obvious use case is converting LLM-generated span coordinates to PDF x,y coordinates for annotations and redactions. This requires using the OCR tokens as the source of truth and generating the text layer from the tokens with consistent preprocessing.
Input Requirements
1. PAWLS Token Layout
The fundamental building-block of our translation layer is the PAWLS token - a format we originally adopted from Allen AI's PAWLS project. Check out detailed typing in types.
PAWLs Token Example:
pawls_tokens = [
{
"page": {"width": 612, "height": 792, "index": 0},
"tokens": [
{"x": 72, "y": 72, "width": 50, "height": 12, "text": "Hello"},
{"x": 130, "y": 72, "width": 50, "height": 12, "text": "World"}
]
}
]
This represents the foundational "source of truth" - the actual positions of text on PDF pages.
2. Derived DataFrames
We then use pandas DataFrames to create efficient indices:
page_df = pd.DataFrame([
{"Page": 0, "Start": 0, "End": 500},
{"Page": 1, "Start": 501, "End": 1000}
])
token_df = pd.DataFrame([
{"Page": 0, "Token_Id": 0, "Char_Start": 0, "Char_End": 5},
{"Page": 0, "Token_Id": 1, "Char_Start": 6, "Char_End": 11}
])
Design Decisions
-
Character-Based Indexing
- Uses character positions as the universal coordinate system
- Makes it trivial to map between text spans and token positions
- Enables precise multi-page span handling
-
Token-First Architecture
- Builds text from tokens rather than trying to match text to tokens
- Guarantees perfect alignment between text and PDF coordinates, so long as you use
doc_text
property provided by PdfDataLayer
to search for spans. - Prevents common OCR/text alignment issues
Why it's Useful
PdfDataLayer
solves several thorny problems:
- Bidirectional Mapping: Seamlessly converts between text positions and PDF coordinates
- Multi-Page Handling: Correctly handles spans that cross page boundaries
- OCR Normalization: Manages common OCR artifacts and character variations
- Efficient Lookups: Uses DataFrame indices for fast position queries
- Clean API: Provides intuitive methods for common operations
This addresses some serious and common real-world document processing challenges and provides a solution that's both powerful and practical.
Basic Usage
1. Importing the Library
Start by importing the necessary components:
from plasmapdf.models.PdfDataLayer import build_translation_layer
from plasmapdf.models.types import TextSpan, SpanAnnotation, PawlsPagePythonType
2. Creating a PdfDataLayer
The core of plasmaPDF is the PdfDataLayer
class. You create an instance of this class using
the build_translation_layer
function:
pawls_tokens: list[PawlsPagePythonType] = [
{
"page": {"width": 612, "height": 792, "index": 0},
"tokens": [
{"x": 72, "y": 72, "width": 50, "height": 12, "text": "Hello"},
{"x": 130, "y": 72, "width": 50, "height": 12, "text": "World"}
]
}
]
pdf_data_layer = build_translation_layer(pawls_tokens)
3. Working with Text Spans
You can extract raw text from a span in the document:
span = TextSpan(id="1", start=0, end=11, text="Hello World")
raw_text = pdf_data_layer.get_raw_text_from_span(span)
print(raw_text)
4. Creating Annotations
To create an annotation:
span_annotation = SpanAnnotation(span=span, annotation_label="GREETING")
oc_annotation = pdf_data_layer.create_opencontract_annotation_from_span(span_annotation)
5. Accessing Document Information
You can access various pieces of information about the document:
print(pdf_data_layer.doc_text)
print(pdf_data_layer.human_friendly_full_text)
print(pdf_data_layer.page_dataframe)
print(pdf_data_layer.tokens_dataframe)
Development Setup
PlasmaPDF uses hatch
for environment and development workflow management. Here's how to get started:
1. Install Hatch
First, install hatch globally:
pip install hatch
2. Development Environment
Hatch automatically manages virtual environments for you. To activate the development environment:
hatch shell dev
3. Running Tests
PlasmaPDF uses pytest for testing. To run tests:
hatch run dev:pytest
For tests with coverage:
hatch run dev:pytest --cov
4. Code Quality Tools
PlasmaPDF comes with several code quality tools configured:
Formatting
To format your code using black
and isort
:
hatch run dev:format
Linting
To run flake8 linting:
hatch run dev:lint
Type Checking
To run mypy type checking:
hatch run types:check
5. Environment Details
PlasmaPDF defines several hatch environments in pyproject.toml
:
dev
: Main development environment with testing and formatting toolstypes
: Environment for type checking with mypy
Each environment has its own dependencies and scripts defined in pyproject.toml
.
6. Code Style
The project follows these standards:
- Line length: 88 characters (Black default)
- Python version: 3.8+
- Strict type checking with mypy
- Black code style
- Isort for import sorting (configured to be compatible with Black)
Advanced Usage
Working with Multi-Page Documents
PlasmaPDF can handle multi-page documents. When you create the PdfDataLayer
, make sure to include tokens for all
pages:
multi_page_pawls_tokens = [
{
"page": {"width": 612, "height": 792, "index": 0},
"tokens": [...]
},
{
"page": {"width": 612, "height": 792, "index": 1},
"tokens": [...]
}
]
pdf_data_layer = build_translation_layer(multi_page_pawls_tokens)
Splitting Spans Across Pages
If you have a span that potentially crosses page boundaries, you can split it:
long_span = TextSpan(id="2", start=0, end=1000, text="...")
page_aware_spans = pdf_data_layer.split_span_on_pages(long_span)
Creating OpenContracts Annotations
To create an annotation in the OpenContracts format:
span = TextSpan(id="3", start=0, end=20, text="Important clause here")
span_annotation = SpanAnnotation(span=span, annotation_label="IMPORTANT_CLAUSE")
oc_annotation = pdf_data_layer.create_opencontract_annotation_from_span(span_annotation)
Utility Functions
PlasmaPDF includes utility functions for working with job results:
from plasmapdf.utils.utils import package_job_results_to_oc_generated_corpus_type
corpus = package_job_results_to_oc_generated_corpus_type(
job_results,
possible_span_labels,
possible_doc_labels,
possible_relationship_labels,
suggested_label_set
)
This function packages job results into the OpenContracts corpus format.
Testing
PlasmaPDF comes with a suite of unit tests. You can run these tests to ensure everything is working correctly:
hatch test
This will run all the tests in the tests
directory.
Conclusion
This quick start guide covers the basics of using PlasmaPDF. For more detailed information, refer to the tests or explore the source code. If you encounter any issues or have questions, please refer to the project's issue tracker or documentation.