RailReader.Renderer.Skia 0.40.0

RailReaderCore

Portable libraries powering RailReader2 and intended for reuse by future companion apps (web, mobile). Distributed as a set of NuGet packages.

Packages

Reference graph

RailReader.Core              ← no native deps, no IO
  ├─ Core.Pdfium             → Core
  ├─ Core.Analysis           → Core
  ├─ Core.Vlm.OpenAI         → Core
  └─ Renderer.Skia           → Core + Core.Pdfium

A future Lite (web/WASM) app would consume RailReader.Core only and substitute its own IPdfService / IPdfTextService / ILayoutAnalyzer / IVlmService implementations (e.g. PDF.js, ORT-Web, browser fetch).

Model-agnostic layout pipeline

Core defines two seams that let any layout-detection model drive RailReader:

• ILayoutAnalyzer — wraps a specific ONNX model and declares its class table + input size + whether it provides reading order via LayoutModelCapabilities. Each detection is stamped with a portable BlockRole; Core never branches on the model-specific class id.
• IReadingOrderResolver — assigns 0..N-1 reading order to detected blocks. Three built-ins ship:
  • ModelOrderResolver (trusts the analyzer's order hints — default pick for models with ProvidesReadingOrder=true)
  • XYCutPlusPlusResolver (column-aware recursive XY-cut, default for non-ordering models — handles two/three-column papers and full-width spanners correctly)
  • TopDownReadingOrderResolver (Y-then-X baseline, retained as a debug fallback)

Core.Analysis ships three analyzers today:

Additional analyzers slot in as further Core.Analysis types or as separate sibling packages — the existing three are the template.

Build & test

dotnet build RailReaderCore.slnx -c Release
dotnet test tests/RailReader.Core.Tests -c Release

Always use -c Release — debug builds are significantly slower on the inference paths.

Models

Core.Analysis consumers need at least one ONNX layout-detection model on disk:

./scripts/download-model.sh           # default — PP-DocLayoutV3 (~50 MB, Apache-2.0)
./scripts/download-model.sh pps       # PP-DocLayout-S (~4.7 MB, Apache-2.0; lightweight)
./scripts/download-model.sh heron     # Docling Heron (~164 MB, Apache-2.0)
./scripts/download-model.sh all       # all three

PP-DocLayout-S is sourced from stefanj0/PP-DocLayout-S-ONNX — a paddle2onnx export of the upstream PaddlePaddle/PP-DocLayout-S checkpoint (no official ONNX exists upstream). Override the source by setting PP_S_ONNX_URL before running the script.

Files land in ./models/. The model search order on disk is defined by Core.Pdfium's LayoutModelLocator (it walks several well-known locations relative to AppContext.BaseDirectory and AppConfig.ConfigDir).

License

MIT — see LICENSE.