Output Structure - Aryn Documentation

Overall Format

JSON Format

The default output format of Aryn DocParse is JSON.

{
  "status": in-progress updates,
  "error": any errors encountered,
  "elements": a list of elements,
  "properties": a list of extracted properties
}

Markdown Format

If the request to Aryn DocParse has the output_format option set to markdown, a successful response will look like this:

{ "status": ...,
  "markdown": "# Title\ndolorem ipsum, quia dolor sit amet consectetur..." }

If group_by_page is set to True in markdown_options, the markdown output will be returned as an array of markdown where each element of the array represents a page:

{
  "status": ...,
  "markdown": ["# Title\ndolorem ipsum", "quia dolor sit", ..., "amet consectetur"]
}

HTML Format

If the request to Aryn DocParse has the output_format option set to html, a successful response will look like this:

{ "status": ...,
  "html": "<h1>Title</h1><p>dolorem ipsum, quia dolor sit amet consectetur...</p>" }

Element Format

General Structure

It is often useful to process different parts of a document separately. For example, you might want to process tables differently than text paragraphs, and typically small chunks of text are embedded separately for vector search. In Aryn DocParse, these chunks are called elements. The Aryn Platform also uses the same format for Documents and Elements. Elements follow the following format:

{"type": type of element (str),
"bbox": Coordinates of bounding box around element (float),
"properties": { "score": confidence score (float),
                "page_number": page number element occurs on (int)},
"text_representation: for elements with associated text (str),
"binary_representation: for Image elements when extract_table_structure is enabled (bytes)}

Each element always has a type, bbox, properties, and text_representation field. The type field indicates the type of the element (e.g., text, image, table, etc.), the bbox field contains the coordinates of the bounding box around the element, the properties field contains additional information about the element (e.g., confidence score, page number, etc.), and the text_representation field contains the text content of the element. The score tag represents the confidence score in the label assignment (Text, Image, Table, etc.). The page_number is the 1-indexed page number the element occurs on. An example element is given below:

{
    "type": "Text",
    "bbox": [
      0.10383546717026654,
      0.31373721036044033,
      0.8960905187270221,
      0.39873851429332385
    ],
    "properties": {
      "score": 0.9369918704032898,
      "page_number": 1
    },
    "text_representation": "It is often useful to process different parts of a document separately. For example, you\nmight want to process tables differently than text paragraphs, and typically small chunks\nof text are embedded separately for vector search. In Aryn DocParse, these\nchunks are called elements.\n"
}

Element Type

"type": one of the element types below

Type	Description
Title	Large Text
Text	Regular Text
Caption	Description of an image or table
Footnote	Small text found near the bottom of the page
Formula	LaTeX or similar mathematical expression
List-item	Part of a list
Page-footer	Small text at bottom of page
Page-header	Small text at top of page
Image	A picture or diagram. When `extract_images` is set to `true`, this element includes a `binary_representation` tag which contains a base64 encoded ppm image file. When `extract_images` is false, the bounding box of the Image is still returned
Section-header	Medium-sized text marking a section
table	A grid of text. See the `extract_table_structure` option to extract information from the table rather than just detecting its presence

Bounding Box

"bbox": coordinates of the bounding box around the element contents

Takes the format [x1, y1, x2, y2] where each coordinate is given as the proportion of how far down or across the screen the element is. For instance, an element that is 100 pixels from the left border of a document 400 pixels wide would have an x1 coordinate of 0.25.

Properties

"properties":
    { "score": confidence between 0 and 1, with 1 being the most confident 
                that this element type and bounding box coordinates are correct (float),
      "page_number": 1-indexed page number the element occurs on (int) }

The score is the model’s “confidence” in its prediction for that particular bounding box. By default, we automatically select bounding boxes to achieve good coverage with high prediction accuracy, but the user can control this by using the threshold parameter (defaults to “auto”). If the user specifies a numeric value between 0 and 1, only Elements with a confidence score higher than the specified threshold value will be kept.

Text Representation

"text_representation": text associated with this element (str)

Text elements contain \n when the text includes a line return.

Binary Representation

When extract_images is set to True, Images include a binary_representation tag which contains a base64 encoded ppm image file of the pdf cropped to the bounds of the detected image. When extract_images is false, the bounding box of the Image is still returned.

"binary_representation": base64 encoded ppm image file of the pdf cropped to the image (bytes)

For a tutorial on how to use the output of Aryn DocParse, see the output walkthrough page.

Table Structure

Tables are represented as a list of cells, where each cell is a dictionary with the following attributes:

{
  "type": "table",
  "bbox": [int, int, int, int],
  "properties": {
    "score": float,
    "title": str,
    "columns": int,
    "rows": int,
    "page_number": int
  },
  "text_representation": str,
  "table": {
    "cells": [
      {
        "content": str,
        "rows": list[int],
        "cols": list[int],
        "is_header": bool,
        "bbox": [int, int, int, int],
        "properties": dict
      }
    ]
  }
}

A table also contains a table tag with individual cells. Each cell has 6 attributes: content, rows, cols, is_header, bbox, and properties. The content attribute contains the text content of the cell, the rows attribute contains the row index of the cell, the cols attribute contains the column index of the cell, the is_header attribute indicates whether the cell is a header cell and is optional, the bbox attribute contains the bounding box of the cell, and the properties attribute contains additional properties of the cell. The text_representation tag contains the text content of the table and is either not specified or is the CSV representation of the table. In the properties dictionary, the title tag represents the string table title if specified, columns is an integer representing the number of columns in the table, and rows is an integer representing the number of rows. They may all not be specified.

Image Structure

{ 
  {
    "type": "Image",
    "bbox": [int, int, int, int],
    "properties": {
      "score": float,
      "image_size": [int, int],
      "image_mode": str,
      "page_number": int
    },
    "binary_representation": bytes
  }
}

If image_extraction_options: {'associate_captions': True} is passed in the request, the output will also include a caption element in the properties dictionary.

"caption": {
  "type": "Caption",
  "bbox": [int, int, int, int],
  "properties": {
    "score": float,
    "_element_index": int,
    "font_size": float
  },
  "text_representation": str,
  "binary_representation": bytes
  }
}

For an image, the binary_representation tag contains a base64 encoded ppm image file of the pdf cropped to the bounds of the detected image. The properties contain extra attributes: image_size, which represents [width, height] of the image, and image_mode, which represents the color mode of the image.

Properties Format

If property_extraction_options: {'schema': [{...}]} is passed in the request, the output will contain a top-level properties element which will contain a dictionary of key value pairs where each key maps to a property in the schema and its value is the result of extraction from the document.

Example schema

{
  "schema": [
    {
      "name": "invoice_number",
      "type": {
        "type": "string",
        "description": "Unique invoice identifier",
        "examples": ["INV-2024-001", "00123", "A-4567"],
        "validators": [
          {
            "type": "boolean_exp",
            "expression": "len() > 0"
          }
        ]
      }
    },
    {
      "name": "invoice_date",
      "type": {
        "type": "date",
        "description": "Date the invoice was issued, if not provided, return null",
        "examples": ["2024-01-15", "2024-03-22"]
      }
    },
    {
      "name": "subtotal",
      "type": {
        "type": "float",
        "description": "Subtotal amount before taxes and fees",
        "examples": [1250.00, 89.99, 2500.50],
        "validators": [
          {
            "type": "boolean_exp",
            "expression": ">= 0"
          }
        ]
      }
    },
    {
      "name": "line_items",
      "type": {
        "type": "array",
        "description": "Array of individual line items on the invoice",
        "item_type": {
          "type": "object",
          "properties": [
            {
              "name": "description",
              "type": {
                "type": "string",
                "description": "Description of the product or service",
                "examples": ["Software Development - 40 hours", "Office Supplies", "Consulting Services"],
                "validators": [
                  {
                    "type": "boolean_exp",
                    "expression": "len() > 0"
                  }
                ]
              }
            },
            {
              "name": "quantity",
              "type": {
                "type": "float",
                "description": "Quantity of items or hours",
                "examples": [1.0, 40.0, 2.5],
                "validators": [
                  {
                    "type": "boolean_exp",
                    "expression": "> 0"
                  }
                ]
              }
            }
          ]
        }
      }
    }
  ]
}

Example output

{
  "properties": {
      "invoice_number": {
        "name": "invoice_number",
        "type": "string",
        "value": "2631",
        "pages": [
          1
        ]
      },
      "invoice_date": {
        "name": "invoice_date",
        "type": "string",
        "value": "2023-03-02",
        "pages": [
          1
        ]
      },
      "subtotal": {
        "name": "subtotal",
        "type": "float",
        "value": 1398.00,
        "pages": [
          1
        ]
      },
      "line_items": {
        "name": "line_items",
        "type": "array",
        "value": [
          {
              "name": null,
              "type": "object",
              "value": {
                "description": {
                  "name": "description",
                  "type": "string",
                  "value": "Appliance:XRS18GGABB CROSLEY 18 TM REF GLASS BL Textured Door",
                  "pages": [
                    1
                  ]
                },
                "quantity": {
                  "name": "quantity",
                  "type": "float",
                  "value": 2.0,
                  "pages": [
                    1
                  ]
                }
              }
          }
        ]
      }
  }
}

Property Metadata

When extracting properties, additional metadata about the extraction will be present under the property_metadata key. This object is a dictionary mapping property names to metadata structures like the following:

{
  "name": "named_insured",
  "type": "string",
  "value": "Global Logistics Solutions Inc.",
  "attribution": {
    "element_indices": [
      4
    ],
    "page": 1,
  },
  "additional_guesses": [
    "Global Logistics Solutions Inc.",
    "Global Logistics Solutions Inc."
  ],
}

The name and type matches what is in the schema, and the value matches what is present in properties. The attribution field contains information about where the extracted value was found — page is the page number in the document, and element_indices is a list of indices of the element containing the property in the elements array. Note that attribution is best effort, so it is possible that attribution will be null even when a property was extracted. For properties of type array, the value field will be an array of metadata objects. The attribution field for all arrays will be null, but the individual array elements will have attribution. If "voting: true" is present in property_extraction_options, then the additional_guesses field will contain the values returned by the other two LLMs used for voting. For a tutorial on how to use the output of Aryn DocParse, see the output walkthrough page.

​Overall Format

​JSON Format

​Markdown Format

​HTML Format

​Element Format

​General Structure

​Element Type

​Bounding Box

​Properties

​Text Representation

​Binary Representation

​Table Structure

​Image Structure

​Properties Format

​Example schema

​Example output

​Property Metadata