Textract
Amazon Textract extracts machine-readable data from scanned documents. It goes beyond optical character recognition: as well as detecting typed and handwritten text, it recovers the structure that gives the text meaning — which value belongs to which form field, which cells belong to which table.
What it extracts
Section titled “What it extracts”- Text detection — typed and handwritten text, returned as words and lines with their positions on the page
- Forms — key/value pairs, so that a printed label and the value written next to it stay associated
- Tables — cell contents grouped by row and column
- Queries — targeted extraction: ask “what is the policy number?” of a document and get the answer rather than the whole page. Custom Queries adapts this to a specific document set.
Specialised APIs
Section titled “Specialised APIs”- AnalyzeExpense — invoices and receipts, returning line items and summary fields such as total and vendor
- AnalyzeID — identity documents including driving licences and passports issued by the US government
- Analyze Lending — mortgage loan packages, routing each page to the appropriate analysis operation and returning per-page or summarised results
Input and processing modes
Section titled “Input and processing modes”Textract accepts image files and PDFs. Synchronous operations handle single-page documents where latency matters; asynchronous operations handle multi-page documents, returning results through a job that is polled or signalled via SNS.
File size and supported languages are fixed quotas. Throughput quotas can be raised through Service Quotas, and the Textract console includes a calculator for estimating what a workload needs.
Typical uses
Section titled “Typical uses”- Building a searchable index over a library of scanned documents
- Feeding structured text into an NLP pipeline, with grouping preserved by line or table cell
- Automating data capture from submitted forms into an existing business workflow
- Classifying incoming documents and routing each type to the right processing path
Practical notes
Section titled “Practical notes”- Results carry confidence scores. Route low-confidence extractions to human review rather than accepting them silently.
- Textract recovers structure, it does not validate meaning. A total that Textract read correctly can still be the wrong total.
- Scanned documents frequently contain personal data. Decide retention for both the source images and the extracted output before building the pipeline.