Skip to main content
Extracts structured fields from PDFs that are too large or too dense for one extraction call. Display name “Large Document Extraction”. Off by default for new agents. It runs the same extraction as Structured Data Extraction, on page chunks of the document. Use it instead of splitting a PDF yourself and extracting each part.

​
Authentication and enablement

No integration. Enable it on the agent. It uses the agent’s extraction model, like Structured Data Extraction.

​
Inputs

It takes the Structured Data Extraction inputs that apply to PDFs:
  • file_url (required): the document.
  • extraction_instructions (required): what to extract.
  • output_schema, model, temperature, seed, reasoning_effort, validation_rules, static_checks, pause_on_validation and disable_metadata.
use_source_citations, use_markdown_ocr, sheets and spreadsheet_agent_mode have no effect here. One input is specific to this tool:
  • max_pages_per_chunk: optional integer, 1 or more. Configured-tool setting only, so the agent cannot change it.
    • Blank: the document is split only when it is too big for one call.
    • Set: no chunk spans more than this many pages, for example 1–5, 6–10, … for a value of 5.
    • When to set it: dense documents with many rows per page, such as long bank statements and transaction reports. They fit one call’s input but write more rows than one call’s output can hold, so rows get lost.

​
How it runs

  1. Convert. TIFF and Office files are converted to PDF first. Any other file type is refused.
  2. Tile. Oversized pages, such as engineering drawing sheets, are cut into tiles so the model can read them.
  3. Plan chunks. Each chunk stays under the model’s input budget:
    • Gemini and Nanonets models: 12 MiB and 800 tiled pages per chunk. The page budget is 200 tiled pages with high media resolution.
    • OpenAI and Anthropic models: 18 MiB and 90 pages per chunk.
    • max_pages_per_chunk, when set, caps chunks further.
  4. Extract. All chunks are extracted at the same time, each with its own text recognition (OCR) for citations. Citation matching then also runs for all chunks at once. A document’s run time is therefore about the time of its slowest chunk.
Each chunk is extracted independently. A chunk never sees the other chunks’ pages, so:
  • A value printed once on page 1, such as a statement date or account holder, appears only in the first chunk’s record.
  • Summary pages are read as ordinary data.
Reconcile across chunks in a later step if your output depends on it.

​
Output

A one-chunk document returns exactly what Structured Data Extraction returns. With more than one chunk, structured_data is a list in page order, with no merging across chunks:
  • A chunk that returns an object becomes one record.
  • A chunk that returns a list contributes each of its items.
  • Every record carries __chunk: {index, start_page, end_page}, with pages numbered in the original document.
  • A chunk that finds nothing contributes no record, so __chunk.index values can skip.
Result metadata adds three counts:
  • chunks_total: chunks planned.
  • chunks_succeeded: chunks that extracted without error.
  • chunks_with_data: chunks that returned at least one value.
page_count is the original page count. The result is published as a ${VAR_N} holding structured_data, so later steps such as a review form or a Python step can read it.

​
Limits and side effects

  • PDF only, after TIFF and Office conversion.
  • Free-tier organisations are limited to the first 100 pages of a file.
  • Billed per page of the original document, at the tier of the extraction model. Nothing is billed when every chunk fails.
  • Many chunks means many model calls at once. A provider rate limit falls back to the fallback model where one is allowed. Otherwise the affected chunks fail.
  • No external writes.

​
Expected errors

  • Missing file_url or extraction_instructions.
  • A file that is not a PDF and cannot be converted to one. The result says so and points to Structured Data Extraction.
  • PDF produced no extractable pages.
  • Failed to extract data: … when every chunk fails.
  • Partial extraction: when some chunks fail, the rest are still returned.
    • Metadata is marked partial_extraction: true with the chunk counts.
    • The result text starts with a warning that those pages are missing.
    • Check for it before treating the output as complete.