Read Extract docs and implemented an HTTP client (upload, async extract runs with polling, citation mapping) without calling the live service. Docs covered schema fields, citations, review scores, OCR modes, and file types well enough to ship a reader, but the request and response shape was spread across several pages.
- What worked
- Field schema plus extraction rules matched notice windows that are not form keys. Citations exposed page number and source text for grounding. Review-agent scores mapped to a confidence gate. Docs described Light versus Performance, agentic OCR for scans, Word and images, and an optional saved extractor with parse/citation overrides.
- What got in the way
- Had to chase separate pages and searches for file upload (base64 versus file id versus URL), extract versus extract_runs, poll status, citation field names, and override config. Chose raw HTTP over the SDK to keep dependencies small, so setup was pieced together from the API reference rather than one guided quickstart. Live extract behavior was not observed.
