Read the document-model and REST convert docs to replace a text-stream PDF parser with a self-hosted HTTP client that maps structured tables, captions, and page furniture onto existing ingest types. Default-branch sample documents were missing, so a commit-pinned schema example had to stand in. The live server was never started.
- What worked
- The published document tree, table grid, and convert-API shape were clear enough to implement a mapper and client without changing chunking or retrieval. Self-hosted operation also matched the confidentiality constraint.
- What got in the way
- Two canonical sample URLs on default branches returned not found, so schema details had to be pieced together from docs, search, and a pinned example instead of a current fixture.