Chunking strategies for PDF documents: page vs paragraph vs heading

Chunking a PDF sounds simple until you try it. Split by page? Some pages are mostly whitespace. Split by paragraph? Some paragraphs are one sentence. Split by heading? Many PDFs have no clear heading structure at all.

Each strategy has tradeoffs. The right choice depends on what your document actually looks like and how your retrieval system uses chunks.

Chunking by page

The simplest approach: each PDF page becomes one chunk (or gets merged into the previous chunk if the page is under the token budget).

Works well for: slide decks, standardized forms, documents where each page covers a distinct topic (like a product catalog with one product per page).

Breaks down when: a single argument or explanation spans multiple pages. The chunk boundary falls in an arbitrary place, and retrieval returns a fragment without setup or conclusion.

Chunking by paragraph

Split on double line breaks. Each paragraph (or group of short paragraphs up to the token budget) becomes a chunk.

Works well for: narrative documents, blog-style content, meeting notes, and anything where individual paragraphs are self-contained statements.

Breaks down when: the document uses short bullet paragraphs. You end up with hundreds of tiny chunks that carry almost no context individually.

Chunking by heading

Detect section boundaries (titles, numbered headings, all-caps lines) and keep each section as a chunk, splitting further only if a section exceeds the token budget.

Works well for: structured reports, policy documents, technical manuals, anything with a table of contents.

Breaks down when: the PDF is a scanned document with no detectable heading structure, or when sections are extremely long (requiring sub-splits that lose the advantage of semantic grouping).

Overlap: the bridge between chunks

Regardless of split strategy, overlap helps. Repeating the last 32 to 64 tokens of chunk N at the start of chunk N+1 means that if relevant information sits right at a boundary, at least one chunk will have it in full.

Too much overlap wastes token budget. Too little means boundary information gets lost. 10% of your chunk size is a practical starting point.

A decision framework

• Document is structured with clear sections? Start with heading mode.
• Document is a narrative or free-form text? Start with paragraph mode.
• Document is a presentation, form, or catalog? Start with page mode.
• Not sure? Try page mode first. It is the safest default because it preserves the original document flow without requiring structural detection.

Testing your chunks

After splitting, read through the first 5 to 10 chunks manually. Ask yourself: if this chunk came back as a retrieval result, would it make sense on its own? If multiple chunks are confusing without the surrounding context, try a larger token budget or a different split mode.

The PDF Chunker lets you switch modes and token budgets without re-uploading, so you can test quickly. Pair it with the Token Counter for a full picture of your document before deciding.