Chunking PDFs with Layout Awareness

PDFs are not just strings of text; they are "containers" of positioned elements. Native character-based chunking often breaks these documents in the middle of a sentence or, worse, in the middle of a multi-line title. Layout-aware chunking attempts to solve this.

Why Standard Chunking Fails PDFs

• Page Breaks: A sentence might start on page 5 and end on page 6. A naive chunker might create a break right in between.
• Headers: Headers are often small strings that carry vital context for the following paragraphs. If separated into their own tiny chunk, they become useless.
• Sidebars: Text in a sidebar might be interleaved with the main body text in a raw extraction, leading to incoherent chunks.

Strategies for Layout-Aware Chunking

1. Header-Based Splitting

Instead of splitting every 500 characters, split whenever a new Heading element is detected.

from langchain.text_splitter import MarkdownHeaderTextSplitter

# If you have converted your PDF to Markdown (using tools like Marker or Nougat)
headers_to_split_on = [
    ("#", "Header 1"),
    ("##", "Header 2"),
]
splitter = MarkdownHeaderTextSplitter(headers_to_split_on=headers_to_split_on)
chunks = splitter.split_text(markdown_text)

2. Boundary Protection

Ensure that a chunk never stops inside a "protected" zone, such as a table or a bulleted list.

3. Page-Level Bucketing

For slide decks or forms, the individual page is often the most semantically cohesive unit.

def chunk_by_page(elements):
    pages = {}
    for element in elements:
        page_num = element.metadata.page_number
        if page_num not in pages:
            pages[page_num] = ""
        pages[page_num] += element.text + "\n"
    return pages

Best Practice: Semantic Grouping

Use an LLM or a specialized model (like LayoutLM) to identify "logical" chunks. These models can see that three disparate paragraphs actually belong to the same "Summary" section based on their visual positioning and style.

Strategy	When to Use	Pros	Cons
Page-by-Page	Presentations, Catalogs	Simple, High Context	Variable Size
Header-to-Header	Whitepapers, Manuals	Logical flow	Requires good parsing
Recursive	General Text	Predictable length	May split topics

Exercises

1. Look at a scientific paper with 2-column layout.
2. Use a "naive" text extractor. Does the text from Column 1 merge correctly with Column 2?
3. How would you design a chunking rule that ensures headers are always kept with at least the first paragraph of their section?