Platform Documentation

Everything you need to understand, use, and extend the PDF Extraction Playground.

← Back to Scan PDF

1. User Guide

The platform allows you to upload a PDF and invoke multiple extraction backends in parallel for qualitative and quantitative comparison.

  1. Open the Scan PDF (home) page.
  2. Drag & drop or browse to select a PDF (full document is always processed).
  3. Select which models to include (Surya, Docling, Mineru).
  4. Click Extract All Pages and wait for processing.
  5. Toggle side‑by‑side comparison to view per‑model outputs and metrics.
Tip:You can re-run extraction with a different model selection without reloading the page—just adjust the checkboxes and click extract again.

2. Technical Notes

Each model produces structured markdown plus annotated page images. Metrics help gauge structural fidelity and performance:

  • Time (ms): Total extraction latency.
  • Blocks: Top-level logical content groupings.
  • OCR Boxes: Count of recognized text regions (if OCR engaged).
  • Chars / Words: Size indicators of produced markdown.
  • Elements: Breakdown (titles, headers, paragraphs, tables, figures).
  • Confidence: Lower values signal fallback or reduced certainty.

Fallback Mode

If native PDF rendering libraries fail to load, a degraded pipeline generates placeholder images and approximate text; confidence values are intentionally lowered. Consult the backend health endpoint for diagnostic detail.

3. Documentation

  • API documentation (auto-generated via FastAPI)
  • User guide explaining how to use the platform
  • Technical documentation covering model selection criteria and limitations

The FastAPI interactive docs are available at /docs (Swagger UI) and /redoc (ReDoc) on the backend service.

4. Configuration & Environment

  • NEXT_PUBLIC_BACKEND_URL: Frontend → backend base URL (default: http://localhost:8000)

5. API Endpoint

POST /api/extract
Content-Type: multipart/form-data
Fields:
  file   - PDF file
  models - comma-separated list: surya,docling,mineru

6. Troubleshooting

  • Uploads fail: Verify backend is running & CORS not blocked.
  • Very slow: Large scans + OCR heavy pages—reduce models.
  • Empty output: Likely fallback mode—inspect backend /health.

7. Roadmap

  • Exportable comparison reports.
  • Advanced diff visualization.
  • Optional page-range filtering.