I used the public parse, extract, and grounding documentation to design an in-region indexing step for referral packets with fax-quality scans, tables, stamps, handwriting, and rotated pages. The docs described page-level boxes, confidence, clinical review routing, and a private or VPC deployment so document bytes stay in a residency region. I encoded an HTTP client from those pages and never called a live deployment. Grounding confidence was described inconsistently, and the pro model suited to poor scans does not emit word-level scores.
- What worked
- The documentation lined up with the intake constraints: vision parsing of low-quality scans, stamps treated as attestation, tables, handwriting, rotation correction before boxes are assigned, and an enterprise container that can run inside a region instead of a public multi-tenant endpoint. That was enough to choose the pro model, post file bytes and markdown inline, and decide which fields need a person to review them.
- What got in the way
- The grounding schema left confidence out while other client notes said atomic grounding can include it. Joining extracted fields back to parse blocks took several separate pages, and there was no Java SDK to follow. Multipart inline uploads, embedding the schema as JSON, and different parse and extract timeouts had to be inferred. Live accuracy, latency, and error behavior were not observed.