autodesk/DocumentAutomation/docs/template-learning.md

Learned document templates

DOCX Workbench supports two deterministic workflows:

1. TemplateLearningService normalizes a .doc or copies a .docx into a temporary working directory, inspects it, adds a safe {{DOC_BODY}} insertion marker to the working copy, and writes a portable .dtemplate ZIP package.
2. TemplateDocumentGenerator opens the package, replaces explicitly supplied scalar tokens, imports a .doc, .docx, or .txt body through the existing main-story importer, and creates DOCX, PDF, generation, and validation outputs.

The package contains manifest.json, template.docx, source metadata, page archetypes, regions, slots, field bindings, mapping suggestions, and a validation profile. It contains no original source binary unless it is the normalized template copy.

All Word automation is late-bound and serialized through the existing STA queue. Binary .doc sources are opened read-only, macros are disabled through Word automation security, and the source SHA-256 is checked before and after normalization. The source body’s headers, footers, section properties, and page setup are excluded by DocxBodyImporter.

The current UI exposes three modes: template learning, document generation, and the existing quick DOC_BODY path. Page archetypes and slot proposals are deterministic suggestions and require later user confirmation; Ollama is not required for the workflow.

Known limitation: the first slice stores the learned structural model and default slots, but does not yet provide a pixel-level page editor or automatic extraction of every scalar field from an arbitrary engineering drawing. Those operations remain explicit user mapping work.