Parsing Workflow

Parsing Workflow

1. Run document_parse_native with the exact granted identifiers.
2. If and only if privacyPolicy is external_vlm_allowed and the format is PDF, run document_analyze_visuals for every page with the exact granted identifiers and policy. Never send local_only content to the VLM.
3. Run document_validate_artifact for the same file version.
4. Run document_read_content and answer the caller's question from that content. Use document_read_artifact only when validation issues or the caller's requested outputs require block/table inspection.
5. Treat remaining warnings as real capability gaps. VLM OCR is probabilistic evidence and does not replace native evidence when native text exists.
6. Return a grounded summary together with the artifact identity, measured counts, validation result, warnings, and quality status. Never infer that a missing stage succeeded.