Installed the Python client, read ProcessRequest, batch types, and entity/page-anchor messages in the package, then implemented a client that converts proto entities into internal field and line objects. Online versus batch request shapes were clear enough to code against a fake backend without a live project.
- What worked
- Install succeeded at a pinned 3.x release. Generated types for documents, entities, normalized values, page anchors, and batch I/O were inspectable and matched the public API closely enough to implement mapping, imageless online calls, and batch split without guesswork on method names.
- What got in the way
- Proto-plus defaults hide a page index of zero, so empty page refs have to be treated as page zero explicitly. Money and date values sit behind oneof accessors and need careful presence checks. Those details were in the types, not called out clearly up front.