Textract

Welcome back, Solutions Architects.
This is Michael Forrester. In this lesson we’ll explore Amazon Textract — a fully managed machine learning service that automatically extracts printed and handwritten text and structured data from scanned documents. Textract converts document content (images, PDFs, handwritten notes) into machine-readable output you can use for analytics, search, or downstream processing. What Textract extracts:

Plain text (printed and handwritten)
Key-value pairs (form extraction)
Tables (with cell relationships)
Signatures (presence and location)
Structured JSON output you can transform into CSV, TXT, or any application-specific format

Textract returns detailed JSON results describing detected text, forms, and tables. You can post-process those JSON results to produce CSV, plain text, or other formats for storage, analytics, or downstream services.

A slide titled "Document Extraction – Categories" showing four colored icons labeled Text Extraction, Form Extraction, Table Extraction, and Signatures. The icons represent different types of data extraction from documents.

Key capabilities and operational notes

Capability	What it does	Typical use cases
Text extraction	Detects text blocks, words, lines, and bounding boxes in the source image	OCR for scanned receipts, invoices, or handwritten notes
Form extraction	Finds key-value pairs and associates form labels with values	Parsing web forms, application forms, tax forms
Table extraction	Reconstructs tables and cell relationships	Exporting invoice line items, financial tables to CSV
Signatures	Detects presence and location of signatures	Document validation and audit workflows
Integration & scale	Returns structured JSON; supports synchronous and asynchronous APIs for single- and multi-page jobs	Serverless pipelines with S3/Lambda, batch processing with async APIs

Operational highlights

Integration: Commonly paired with Amazon S3, AWS Lambda, Amazon Comprehend, and Amazon DynamoDB to build an end-to-end pipeline.
Scalability: Use asynchronous APIs (Start/Get job pattern) for multi-page or high-volume workloads to avoid Lambda timeouts.
Compliance: Textract is HIPAA-eligible. Use encryption at rest (S3), secure IAM roles (least privilege), and VPC endpoints as needed.

I’ve used Textract on poor handwriting and found it produced usable results with limited manual correction. Pairing Textract output with NLP (Amazon Comprehend or custom models) improves entity extraction and PII detection. Common architecture pattern

Documents are uploaded to Amazon S3.
An S3 event triggers an AWS Lambda function.
The Lambda function calls Amazon Textract (synchronous for quick single-page jobs; asynchronous for multi-page or large batches).
Textract returns structured JSON. The Lambda stores JSON in S3, indexes parsed fields into DynamoDB, or forwards extracted text to Amazon Comprehend for NLP.

You can run Textract from the console for quick experiments (upload an image or PDF and inspect the JSON), or embed it in serverless pipelines for automated, high-volume processing.

An architecture diagram titled "Textract" showing data flow from Amazon S3 to AWS Lambda to Amazon Textract and then to Amazon DynamoDB. Each AWS service is represented by its icon and connected with arrows.

Quick examples

Synchronous (single image / quick test) — AWS CLI:

aws textract detect-document-text --document '{"S3Object":{"Bucket":"my-bucket","Name":"page1.png"}}' --region us-east-1

Asynchronous (multi-page PDF / batch processing) — start job and then get results:

# Start async job
aws textract start-document-text-detection --document-location '{"S3Object":{"Bucket":"my-bucket","Name":"document.pdf"}}' --region us-east-1

# Retrieve results (use JobId from start response)
aws textract get-document-text-detection --job-id <JOB_ID> --region us-east-1

Sample (simplified) JSON output structure

{
  "DocumentMetadata": { "Pages": 2 },
  "Blocks": [
    { "BlockType": "PAGE", "Geometry": { "BoundingBox": { ... } } },
    { "BlockType": "LINE", "Text": "Invoice #1234", "Geometry": { ... } },
    { "BlockType": "KEY_VALUE_SET", "EntityTypes": ["KEY"], "Relationships": [ ... ] }
  ]
}

Practical tips and best practices

Use asynchronous APIs for large or multi-page documents to avoid Lambda time limits.
Persist Textract JSON to S3 for auditing, replayability, and long-term storage.
Store parsed key fields (IDs, dates, totals) in DynamoDB for fast queries and indexing.
Combine Textract output with Amazon Comprehend or custom NLP models to extract entities, detect PII, or perform sentiment analysis.
Secure the pipeline: enable S3 encryption, enforce IAM least-privilege, and consider VPC endpoints for private access.

When processing sensitive data (PII or PHI), ensure your AWS account and pipeline meet compliance requirements. Use encryption, strict IAM policies, and logging/monitoring to maintain an audit trail.

Recommended integrations and references

Amazon S3 — storage for source documents and Textract output
AWS Lambda — orchestration and lightweight processing
Amazon DynamoDB — fast lookup and indexing of parsed fields
Amazon Comprehend — NLP for entity extraction, PII detection, and sentiment
AWS Identity and Access Management (IAM) — secure permissions for Textract and downstream services

Additional resources

That’s AWS Textract — a managed, scalable solution for extracting text and structured data from documents and turning them into usable digital information. Michael Forrester — catch us on Slack if you have any questions.

Introduction

Services - Networking

Services - Storage

Services - Compute

Services - Database

Services - Application Integration

Services - Data and ML

Services - Migration and Transfer

Services - Management and Governance

Services - Security

Bringing it all together

Designing for Security

Designing for Reliability

Designing for Performance

Designing for Cost-Optimization

Applying your Design Skills

Watch Video