Classifies an inbound document, files it automatically when the model is confident, and routes it to a human reviewer when it is not. VisML 2026-08-04T09:00:00Z Apache-2.0 ai orchestration human-in-the-loop Vision-capable model used for both OCR and classification. Boundary node. Emits one document per invocation.

Raw uploaded bytes.

Pure function of the document bytes, so it is freely retryable and its result is cacheable by digest.

Returns a category from the taxonomy plus the model's own confidence. The confidence is what the next node branches on.

Cases are evaluated in document order; the first true case wins.

Blocks until a reviewer decides. idempotent="false" stops a runtime from silently enqueueing a second review on retry.

Reached when the model is not confident enough to be worth a reviewer's time. A terminal, deliberate outcome — not an error. Target of the error edge from classify. Reached only after every retry is exhausted. Traversed only when classify reaches FAILED after retries.