Intelligent Document Processing at Scale with Generative AI and Amazon Bedrock Data Automation

Pulling structured fields out of contracts, emails and reports looks solved until you try it. Named entity recognition, the classic approach, only handles text, only finds the categories in its training set, and has nothing useful to say when the field you want is a sentiment score or a one-line summary. Generative models remove that ceiling without annotation or fine-tuning. The application we built takes a set of documents and a list of attributes and gives you back a table, with Amazon Bedrock Data Automation doing the extraction and handling parsing, context management and model selection behind a single API.

For most pipelines of this kind that is the right default, but not for all. Some organizations must run self-hosted models for regulatory reasons, some builder teams want control over every step, and Bedrock Data Automation was only available in us-west-2 and us-east-1 as of June 2025. So the pipeline keeps two other paths open, calling Bedrock models directly or running Amazon Textract for OCR and passing the text to a model, and you choose per run.

Around all of that sits fairly ordinary infrastructure, which in practice is where most of the work goes. Step Functions orchestrates the pipeline and fans out across documents with a Map state, Lambda makes the calls, S3 holds inputs and results, and a container on ECS behind CloudFront gives non-technical users a UI, with Cognito in front. It deploys as CDK, and the code is on GitHub under the MIT license.