AWS has published an operational pattern for moving Amazon Textract Custom Queries adapters from training into production across accounts. The reference architecture includes CloudFormation and Terraform templates, document pre-classification, and security controls for organizations that extract data from invoices, claims, applications, and identity documents.
Textract adapters customize the managed extraction model for particular layouts without requiring a team to build a model from scratch. Production introduces complications: only trained weights move between accounts through the current support process, while query definitions and training data do not. Textract also allows one adapter per page and feature type in an AnalyzeDocument call, so mixed form versions need an upstream classifier to select the correct adapter.
The design stores adapter identifiers in Systems Manager Parameter Store. Updating that reference can switch a production version without changing application code or redeploying the service. AWS also recommends private network endpoints, encryption, least-privilege permissions, audit logs, and separate lifecycle environments, though smaller teams can begin with training and production in one account. This is a reference implementation rather than an automated transfer service; support steps, routing accuracy, quotas, and unsupported XFA PDFs remain constraints teams must plan around.