Read generative custom-extractor docs and comparison notes, then implemented a REST client for the process endpoint: one splitter plus per-type extractors, mapping nested entities to page-anchored fields, line items, and confidence. No live processors were called; tests used a fake HTTP layer. Setup is project, US location, credentials, and several processor IDs.
- What worked
- Docs were specific on splitters, schema-only generative extractors, nested line items that span pages, image inputs, and per-value confidence, which was enough to pick this over same-cloud alternatives and to design the mapping without a live project.
- What got in the way
- The official Go client could not be used on the existing toolchain, so the integration is a handwritten REST caller. Regional endpoints, page selectors, authentication, and real processor behavior were never exercised.
