View a markdown version of this page

Document summarization and figure captioning - Amazon Textract

Document summarization and figure captioning

Amazon Textract Insights can generate a natural-language summary of a document and captions for the figures it contains. You enable these through the project's generative configuration, and they appear in the extraction results.

Enabling generative output

You configure generative output on the project through GenerativeConfiguration, which has two independent capabilities:

  • Summarization – Generates a summary of the document. If your document includes several document types, Amazon Textract Insights creates a separate summary for each logical sub-document.

  • FigureCaption – Generates captions for the figures detected in the document. By default, captioning is limited to five figures per page.

You enable the generative configuration when you create or update a project, alongside the extraction configuration. Each capability you turn on adds processing work, so enable only the ones you need. See Working with projects.

aws textract create-project \ --project-name "summaries-and-captions" \ --extraction-configuration '{ "BoundingBox": { "State": "ENABLED" }, "Granularity": { "Types": ["DOCUMENT", "ELEMENT"] }, "OutputFormat": {"TextFormat": {"Types": ["MARKDOWN"]}, "AdditionalFileFormat": {"State": "DISABLED"}} }' \ --generative-configuration '{ "Summarization": { "State": "ENABLED" }, "FigureCaption": { "State": "ENABLED" } }' \ --region us-east-1

To turn either capability off later, call update-project with the full GenerativeConfiguration, setting that capability's State to DISABLED. Omitting --generative-configuration on create-project leaves all generative output disabled.

Document summaries

When summarization is enabled, the document member of the extraction result.json includes two fields:

  • summary – A natural-language summary of the document's content, typically a few sentences.

  • description – A single sentence that names the document type and its subject, for example "This is an invoice about goods sold."

"document": { "description": "This is an invoice about goods sold.", "summary": "This document is an invoice from one company to another, dated earlier this year. It lists the products sold and their unit prices, along with the invoice total.", "statistics": { "...": "..." } }

If summarization is not enabled, both summary and description are absent.

Use the summary to give downstream consumers — a reviewer, a search index, or an LLM prompt — a quick orientation to the document without reading the full extracted text.

Figure captions

When figure captioning is enabled, Amazon Textract Insights writes a descriptive caption for each figure it detects, such as a chart, graph, or image. Where the caption appears depends on the granularity you request:

  • With ELEMENT granularity, each figure is an element of type FIGURE with a sub_type (for example CHART), a title taken from the document, and a summary that holds the generated caption. The caption is also rendered into the element's representation, for example as Markdown image alt text. statistics.figure_count reports how many figures were found.

  • Without ELEMENT granularity (for example PAGE only), there is no elements section, but the caption is still embedded in the page text representation at the figure's position.

{ "type": "FIGURE", "sub_type": "CHART", "title": "FIG. 3. Comparison between the experimental group and control over 28 days.", "summary": "Title: Comparison of Bone Healing Between Experimental and Control Groups Over 28 Days\nThe figure presents a comparison between the experimental group and the control group...", "reading_order": 9, "locations": [ { "page_index": 0, "bounding_box": { "left": 0.06, "top": 0.76, "width": 0.43, "height": 0.19 } } ], "representation": { "markdown": "![Title: Comparison of Bone Healing Between Experimental and Control Groups Over 28 Days\n..." } }

A document with no figures has a figure_count of 0 and produces no captions.

Enabling figure captioning limits the maximum number of pages supported per document. If you enable figure captioning on a longer document, the job fails with ErrorType DocumentTooLargeException. Summarization has no separate page limit. For these and other limits, see Limits and quotas.

Relationship to granularity

Generative output appears in the extraction result, so the sections it populates depend on the granularity you configure:

  • The document summary appears in the document member. Include DOCUMENT in Granularity.Types to receive the summary together with the whole-document text.

  • Captions appear as FIGURE elements when you request ELEMENT granularity, and inline in the page text otherwise.

See Extraction results for the full structure and Output formats for LLM and RAG workflows for the text formats the summary and captions are rendered in.

Next steps