Document AI fails in regulated contexts when the first diagram points at a public multimodal API. The alternative is a pipeline you operate: ingest, OCR, structure, analyze — with models that never leave your perimeter.
Split OCR workers from product APIs. Use vision models to fix layout, not only to chat about pages. Keep family or case scoping in the application layer so a clever model cannot become an access-control bypass.
OCR is infrastructure
Paddle-style OCR with bounding boxes gives you auditability. Dumping whole-page text into a chat model hides errors.
GPU Compose stacks make this practical — if you document NVIDIA toolkit assumptions honestly.
Private LLMs summarize; they do not authorize
Ollama (or equivalent) is for timelines and insights after retrieval authorization succeeds.
Store originals and derived markdown with lineage so clinicians or families can trust the trail.
Takeaways
- Local document AI is a pipeline, not a chatbot.
- Separate OCR, app API, and LLM runtimes.
- Authorization stays in your application, always.