Aws Textract

Textract goes beyond basic OCR — it understands document structure, extracting tables with cell relationships, form fields with key-value pairing, and specific document types (invoices, receipts, identity documents) with specialized analysis features. For agents in document processing pipelines: sync API for single pages, async API for multi-page PDFs, AnalyzeExpense for invoice data extraction, AnalyzeID for identity document processing. S3-native for large files. Confidence is docs-derived.

AWS SDK (boto3): tx = boto3.client("textract"). Sync (single page): tx.detect_document_text(Document={Bytes: image_bytes}) or S3Object. tx.analyze_document(Document=..., FeatureTypes=["TABLES", "FORMS"]) extracts tables and forms. tx.analyze_expense(Document=...) for invoices/receipts. Async (multi-page): tx.start_document_analysis(DocumentLocation={S3Object: {Bucket, Name}}, FeatureTypes=["TABLES"]). tx.get_document_analysis(JobId=job_id) polls. S3 integration: upload file, pass S3 reference.

IAM auth: AWS credentials. Per-operation: textract:DetectDocumentText, textract:AnalyzeDocument, textract:AnalyzeExpense, textract:AnalyzeID, textract:StartDocumentAnalysis, textract:GetDocumentAnalysis. S3 access for async: textract:GetObject on S3 bucket. Managed identity for EC2/Lambda. HTTPS enforced.

InvalidS3ObjectException for inaccessible S3 objects. UnsupportedDocumentException for unsupported file types. Async job states: IN_PROGRESS → SUCCEEDED/FAILED. Poll with exponential backoff. Textract SLA: 99.9%. Accuracy high for printed text; lower for handwriting. Page limits: sync up to 1 page (JPEG/PNG) or 1-page PDF; async up to 3000 pages.

docs.aws.amazon.com/textract covers AnalyzeDocument, expense/ID features, async workflow, and response parsing. Getting started: boto3 + image, first detect_document_text in under 5 minutes. Free tier: 1000 pages/month for first 3 months. Textract Response Parser library (trp) for structured output. Community via AWS Forums.