spaCy Layout: Process PDFs, Word documents and more with spaCy

This plugin integrates with Docling to bring structured processing of PDFs, Word documents and other input formats to your spaCy pipeline. It outputs clean, structured data in a text-based format and creates spaCy's familiar Doc objects that let you access labelled text spans like sections or headings, and tables with their data converted to a pandas.DataFrame.

This workflow makes it easy to apply powerful NLP techniques to your documents, including linguistic analysis, named entity recognition, text classification and more. It's also great for implementing chunking for RAG pipelines.

📖 Blog post: "From PDFs to AI-ready structured data: a deep dive" – A new modular workflow for converting PDFs and similar documents to structured data, featuring spacy-layout and Docling.

📝 Usage

⚠️ This package requires Python 3.10 or above.

pip install spacy-layout

After initializing the spaCyLayout preprocessor with an nlp object for tokenization, you can call it on a document path to convert it to structured data. The resulting Doc object includes layout spans that map into the original raw text and expose various attributes, including the content type and layout features.

import spacy
from spacy_layout import spaCyLayout

nlp = spacy.blank("en")
layout = spaCyLayout(nlp)

# Process a document and create a spaCy Doc object
doc = layout("./starcraft.pdf")

# The text-based contents of the document
print(doc.text)
# Document layout including pages and page sizes
print(doc._.layout)
# Tables in the document and their extracted data
print(doc._.tables)
# Markdown representation of the document
print(doc._.markdown)

# Layout spans for different sections
for span in doc.spans["layout"]:
    # Document section and token and character offsets into the text
    print(span.text, span.start, span.end, span.start_char, span.end_char)
    # Section type, e.g. "text", "title", "section_header" etc.
    print(span.label_)
    # Layout features of the section, including bounding box
    print(span._.layout)
    # Closest heading to the span (accuracy depends on document structure)
    print(span._.heading)

If you need to process larger volumes of documents at scale, you can use the spaCyLayout.pipe method, which takes an iterable of paths or bytes instead and yields Doc objects:

paths = ["one.pdf", "two.pdf", "three.pdf", ...]
for doc in layout.pipe(paths):
    print(doc._.layout)

spaCy also allows you to call the nlp object on an already created Doc, so you can easily apply a pipeline of components for linguistic analysis or named entity recognition, use rule-based matching or anything else you can do with spaCy.

# Load the transformer-based English pipeline
# Installation: python -m spacy download en_core_web_trf
nlp = spacy.load("en_core_web_trf")
layout = spaCyLayout(nlp)

doc = layout("./starcraft.pdf")
# Apply the pipeline to access POS tags, dependencies, entities etc.
doc = nlp(doc)

Tables and tabular data

Tables are included in the layout spans with the label "table" and under the shortcut Doc._.tables. They expose a layout extension attribute, as well as an attribute data, which includes the tabular data converted to a pandas.DataFrame.

for table in doc._.tables:
    # Token position and bounding box
    print(table.start, table.end, table._.layout)
    # pandas.DataFrame of contents
    print(table._.data)

By default, the span text is a placeholder TABLE, but you can customize how a table is rendered by providing a display_table callback to spaCyLayout, which receives the pandas.DataFrame of the data. This allows you to include the table figures in the document text and use them later on, e.g. during information extraction with a trained named entity recognizer or text classifier.

def display_table(df: pd.DataFrame) -> str:
    return f"Table with columns: {', '.join(df.columns.tolist())}"

layout = spaCyLayout(nlp, display_table=display_table)

Serialization

After you've processed the documents, you can serialize the structured Doc objects in spaCy's efficient binary format, so you don't have to re-run the resource-intensive conversion.

from spacy.tokens import DocBin

docs = layout.pipe(["one.pdf", "two.pdf", "three.pdf"])
doc_bin = DocBin(docs=docs, store_user_data=True)
doc_bin.to_disk("./file.spacy")

⚠️ Note on deserializing with extension attributes: The custom extension attributes like Doc._.layout are currently registered when spaCyLayout is initialized. So if you're loading back Doc objects with layout information from a binary file, you'll need to initialize it so the custom attributes can be repopulated. We're planning on making this more elegant in an upcoming version.

+ layout = spacyLayout(nlp)
doc_bin = DocBin(store_user_data=True).from_disk("./file.spacy")
docs = list(doc_bin.get_docs(nlp.vocab))

🎛️ API

Data and extension attributes

layout = spaCyLayout(nlp)
doc = layout("./starcraft.pdf")
print(doc._.layout)
for span in doc.spans["layout"]:
    print(span.label_, span._.layout)

dataclass PageLayout

dataclass DocLayout

dataclass SpanLayout

class spaCyLayout

method spaCyLayout.__init__

Initialize the document processor.

nlp = spacy.blank("en")
layout = spaCyLayout(nlp)

method spaCyLayout.__call__

Process a document and create a spaCy Doc object containing the text content and layout spans, available via Doc.spans["layout"] by default.

layout = spaCyLayout(nlp)
doc = layout("./starcraft.pdf")

method spaCyLayout.pipe

Process multiple documents and create spaCy Doc objects. You should use this method if you're processing larger volumes of documents at scale. The behavior of as_tuples works like it does in spaCy's Language.pipe.

layout = spaCyLayout(nlp)
paths = ["one.pdf", "two.pdf", "three.pdf", ...]
docs = layout.pipe(paths)

sources = [("one.pdf", {"id": 1}), ("two.pdf", {"id": 2})]
for doc, context in layout.pipe(sources, as_tuples=True):
    ...

💡 Examples and code snippets

This section includes further examples of what you can do with spacy-layout. If you have an example that could be a good fit, feel free to submit a pull request!

Visualize a page and bounding boxes with matplotlib

import pypdfium2 as pdfium
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
import spacy
from spacy_layout import spaCyLayout

DOCUMENT_PATH = "./document.pdf"

# Load and convert the PDF page to an image
pdf = pdfium.PdfDocument(DOCUMENT_PATH)
page_image = pdf[2].render(scale=1)  # get page 3 (index 2)
numpy_array = page_image.to_numpy()
# Process document with spaCy
nlp = spacy.blank("en")
layout = spaCyLayout(nlp)
doc = layout(DOCUMENT_PATH)

# Get page 3 layout and sections
page = doc._.pages[2]
page_layout = doc._.layout.pages[2]
# Create figure and axis with page dimensions
fig, ax = plt.subplots(figsize=(12, 16))
# Display the PDF image
ax.imshow(numpy_array)
# Add rectangles for each section's bounding box
for section in page[1]:
    # Create rectangle patch
    rect = Rectangle(
        (section._.layout.x, section._.layout.y),
        section._.layout.width,
        section._.layout.height,
        fill=False,
        color="blue",
        linewidth=1,
        alpha=0.5
    )
    ax.add_patch(rect)
    # Add text label at top of box
    ax.text(
        section._.layout.x,
        section._.layout.y,
        section.label_,
        fontsize=8,
        color="red",
        verticalalignment="bottom"
    )

ax.axis("off")  # hide axes
plt.show()