Expand description
Pluggable vision backend — the ML boundary.
The core is deterministic and has zero ML dependencies. Everything that
needs a model — OCR for scanned pages, structure recognition for borderless
tables — goes through this trait, not the core. The default NoopBackend
does no inference: a scanned page is surfaced as a NeedsOcr warning and left
for a real backend (an ONNX/Tesseract crate, or a Python-side cloud OCR) to
fill in. Reference backends live in separate optional crates so the core never
links a model runtime.
Structs§
- Noop
Backend - The default backend: no model inference. Scanned pages surface as warnings, keeping the core free of any ML runtime.
Enums§
- Backend
Error - An error from a vision backend.
Traits§
- Vision
Backend - A backend that can recover content the deterministic core cannot.