API reference

One image, three supported bean-leaf classes, an accepted or abstained decision.

POST /v1/predictions

Send a multipart field named file containing JPEG, PNG, or WebP. Default maximum file size: 5 MiB. Filenames are not used for inference.

curl --fail-with-body -F 'file=@leaf.jpg;type=image/jpeg' http://127.0.0.1:8000/v1/predictions

HTTP 200 includes schema version, decision, prediction (only when accepted), class scores, confidence, margin, normalized entropy, quality findings, policy thresholds, model revision, and request ID.

HTTP 413: upload too large. HTTP 415: unsupported type. HTTP 422: invalid input. HTTP 503: model unavailable. HTTP 500: internal or malformed model output.

Operations

• GET /health/live: process liveness; does not load weights.
• GET /health/ready: 503 until a successful first inference loads the model; then 200.
• GET /metrics: Prometheus prediction outcomes and latency.
• GET /openapi.json: complete machine-readable response schemas.

Educational decision-support demo; this is not a field-validated diagnosis.