FinBank receives multi-page PDF invoices (3–10 pages each) in an S3 bucket and needs to extract structured key/value fields (InvoiceNumber, InvoiceDate, InvoiceTotal) and persist the results to DynamoDB. Several teams tried DetectDocumentText but the key/value pairing is unreliable. What is the most appropriate Textract approach and configuration to reliably extract structured form fields at scale in this AWS environment?

The correct approach leverages StartDocumentAnalysis with FORMS feature type, which is designed for extracting structured key/value pairs from documents like invoices. Since the invoices are multi-page and processing needs to be at scale, the asynchronous StartDocumentAnalysis is appropriate. Parsing KEYVALUESET blocks and their RELATIONSHIPS allows for accurate mapping of keys to their corresponding values. The service role permissions are crucial for Textract to access S3. Synchronous AnalyzeDocument is not suitable for multi-page documents at scale and TABLES feature type is for tabular data, not key/value pairs. Amazon Comprehend's DetectEntities is for general entity recognition, not structured key/value extraction from forms. AnalyzeExpense is a specialized API for expense receipts, not general invoices, and may not cover all required fields or be the most cost-effective for this specific use case.