Parse Outputs

​

High-Level Structure

{
  "task_id": "8b7e7e8a-...",
  "status": "Succeeded",
  "output": {
    "file_name": "document.pdf",
    "page_count": 2,
    "chunks": [
      // ... array of chunk objects
    ],
    "pages": [
      // ... array of page objects with full-page images
    ]
  },
  // ... other task metadata
}

​

Chunks and Segments

• Segments: The smallest building blocks. Each segment corresponds to a single, identified element from the source document.
• Chunks: A logical grouping of one or more segments. Each chunk includes chunk_length, content, embed, and segments[]. For RAG applications, chunks are the units of information that are typically embedded and retrieved.

​

Key Output Fields

​

1. content

• Tables: Converted to HTML to maintain complex col/row-span structure.
• Images: Converted to a robust markdown description, with charts/graphs including a tabular representation.
• Forms: Converted to HTML for structured key-value representation.
• Legends: Converted into a key-value markdown table.
• Formulas: Converted to LaTeX strings for perfect mathematical representation. They can even be embedded within an HTML table if a formula appears inside a cell.
• GraphicalItems: Simple OCR extraction of any text present.
• Text-type: Text-heavy segments like title, section headers, list-items and text blocks are converted into markdown.

{
  "segment_type": "Table",
  "content": "<table><tr><td>The formula is:</td><td>\\( E=mc^2 \\)</td></tr></table>"
}

​

2. embed

• It includes the content and, if present, the description. This helps optimize the table segments without contaminating the content field.
• This is the field used for calculating token counts when chunking, ensuring chunks fit your target length.

{
  "chunk_id": "chunk-1-...",
  "chunk_length": 45,
  "embed": "The table shows a 15% increase in Q2 revenue for Widget A... | Product | Q1 | Q2 | ...",
  "segments": [ /* ... */ ]
}

​

3. bbox (Bounding Box)

• The coordinates (left, top, width, height) are pixel (px) values in the page coordinate space. For resolution‑independent rendering, normalize them using the page dimensions — for example, left_pct = left / page_width, top_pct = top / page_height, and likewise for width/height. Use page_width/page_height on the segment (or the page’s pg_width/pg_height).
• The dpi in the pages array describes the pixel resolution of the rendered page image. You do not need dpi when using normalized percentages; it is helpful only when mapping directly to a specific raster image in pixels or when generating images at a different scale.

{
  "segment_type": "Text",
  "bbox": { "left": 100, "top": 250, "width": 500, "height": 50 },
  "page_number": 1,
  "page_height": 1584,
  "page_width": 1224,
  ...
}

​

Spreadsheet-Specific Outputs (ss_*) Preview

A financial spreadsheet showing intelligent segmentation of tables and charts.

​

Key Fields

• ss_range: The cell range for the segment in A1 notation (e.g., A1:D10).
• ss_cells: A detailed array of each cell in the segment, including its original formula, value, text and styling. Allows you to see both the raw formula (=SUM(B2:B10)) and its calculated result ($55,000).
• ss_header_*: Fields identifying the detected header for a table, such as ss_header_range, ss_header_text, ss_header_bbox, and ss_header_ocr. Headers are intelligently associated even if they are not directly adjacent to the table.

• Create Interactive Experiences: Use ss_range to build native citation experiences that let users click data and jump to the precise source cell in a viewer.
• Get Cleaner LLM Context: Combine layout analysis with precise cell data to identify tables, associate headers, and filter out irrelevant cells. This provides cleaner, more meaningful context for LLM processing.
• Build Powerful Spreadsheet Agents: Use the ss_* fields to build AI agents that can read, analyze, and even write back to spreadsheets. Understanding cell formulas and values enables agents to automate tasks like updating financial models, correcting entries, or adding new rows.