Chunk Types

​

Chunk Definition

A chunk is a discrete element extracted from a document, such as a block of text, a table, or a figure.

​

Chunk Overview

When you send a document to the API, it analyzes the content on each page, breaks it down into meaningful elements, and returns each one as a chunk. Each chunk includes structured data that describes the content of the chunk and the location of the chunk in the document. This structure makes it easier to understand the extracted data and use it for downstream tasks. Extracted chunks are included in the API response.

​

Semantic Chunking

The API uses semantic chunking, which means it intelligently groups content based on meaning rather than just layout or formatting. Instead of splitting documents at arbitrary points like fixed lengths or paragraph breaks, the API identifies coherent units of information (like complete ideas, logical sections, or related data) and extracts them as individual chunks. Semantic chunking improves the relevance and usability of the extracted content, especially in downstream tasks like search, retrieval, and analysis.

​

Why Do We Create Chunks?

Chunking makes downstream tasks faster, more accurate, and easier to scale. It serves several key purposes:

• Enables downstream apps to process large documents efficiently: Chunking allows applications like RAG systems and LLMs to index and retrieve smaller, meaningful segments instead of full documents. This helps avoid input size constraints, such as token limits.
• Improves retrieval granularity: Smaller, semantically meaningful units allow for more accurate and relevant results in downstream tasks like question answering and summarization.
• Supports downstream semantic search and embeddings: Well-structured chunks provide better inputs for embedding and make it easier to index and retrieve information during search.
• Maintains human readability: Chunking reflects how a human would naturally read the document, maintaining the visual and logical relationships between elements on the page.

​

Chunk Types

Each chunk is labeled with a chunk type (chunk_type or type, depending on the API used), which identifies what kind of content it represents. The chunk types returned by are:

• text
• table
• marginalia
• figure
• logo: This is only available when using
• card: This is only available when using
• attestation: This is only available when using
• scan_code: This is only available when using

​

Text

A text chunk type is an element that consists entirely of characters (letters and numbers), such as:

• paragraphs
• titles and headings
• lists
• form fields
• checkboxes
• radio buttons
• equations
• code blocks
• handwritten text

​

Output for Key-Value Pairs

If the text content has key-value pairs, like form fields, the extracted data will be returned as key-value pairs separated by line breaks (\n).

​

Example: Paragraph

Here is an example of the API marking a paragraph as a text chunk: Here is the rendered Markdown for that chunk:

​

Example: Lists

Here is an example of the API marking a list as a text chunk: Here is the rendered Markdown for that chunk:

​

Table

A table chunk type is a grid of rows and columns containing data. doesn’t require gridlines to be present, and typically interprets well-aligned sets of data to be part of a table. For example, part of a receipt can be extracted as a table if the purchased items align with the costs.

​

Example: Receipt

Here is an example of the API marking receipt line items as a table chunk: Here is the rendered Markdown for that chunk:

​

Example: Earnings Statement

Here is an example of the API marking part of an earnings statement as a table chunk: Here is the rendered Markdown for that chunk:

​

Marginalia

A marginalia chunk type is a set of text in the top, bottom, or side margins of a document, including:

• page headers
• page footers
• page numbers
• handwritten notes in margins
• line numbers on one side of a page

​

Example: Header and Page Number

Here is an example of the API marking a header and page number as a page_header chunk: Here is the rendered Markdown for that chunk:

​

Figure

A figure chunk type is an element that contains visual or graphical non-text content, including:

• pictures
• graphs (bar graphs, line graphs, etc.)
• flowcharts
• diagrams

If using , the figure chunk type also includes:

• logos
• QR codes
• barcodes
• stamps
• signatures
• ID cards

​

Example: Medical Imaging

Here is an example of the API marking a pathology image as a figure chunk: Here is the rendered Markdown for that chunk:

​

Example: Bar Chart

Here is an example of the API marking a bar chart as a figure chunk: Here is the rendered Markdown for that chunk:

​

Logo

A logo chunk type identifies logos.

​

Example: Logo in Header

Here is an example of the API marking a logo in a document header as a logo chunk: Here is the rendered Markdown for that chunk:

​

Card

A card chunk type identifies:

• ID cards
• driver licenses

​

Example: Driver’s License

Here is an example of the API marking a driver’s license as a card chunk: Here is the rendered Markdown for that chunk:

​

Attestation

An attestation chunk type includes:

• signatures
• stamps
• seals

​

Example: Signature

Here is an example of the API marking a signature as an attestation chunk: Here is the rendered Markdown for that chunk:

​

Scan_code

A scan_code chunk type identifies:

• QR codes
• bar codes

​

Example: Bar Codes

Here is an example of the API marking two barcodes as scan_code chunks: Here is the rendered Markdown for these chunks:

​

Deprecated Chunk Types

Some chunk types were deprecated and consolidated into other types. These changes were introduced in the library v0.2.1, and were rolled out to the API on Thursday, May 22, 2025. These chunk types were consolidated into marginalia:

• page_header
• page_footer
• page_number

These chunk types were consolidated into text:

• title
• form
• key_value

​

Action Required When Using Library

If you use the library and your scripts or workflows use any of the deprecated chunk types, update your code to use the new types. How the library handles the deprecated chunk types depends on the version you’re using:

• Upgrade to v0.2.1 to use the new chunk types.
• If using v0.0.13 to v​​0.1.3, the marginalia type doesn’t exist and will fallback to page_header.
• If using v0.0.12 or earlier, the code no longer works.

​

Action Required When Calling the API Directly

If you call the API directly and your scripts or workflows use any of the deprecated chunk types, update your code to use the new types. We made these same changes (consolidating the chunk types) to the API on Thursday, May 22. Starting May 22, the API stopped using the deprecated types in the response. If your code uses the deprecated chunk types, the code will no longer work.