OCR/testing_strategy.md

Testing Strategy: Template Mapping Engine

This document outlines the testing strategy for the Template Mapping Engine to ensure production readiness.

1. Backend Testing (Python/FastAPI)

Unit Testing

• Framework: pytest
• Database: Use an in-memory SQLite database (sqlite:///:memory:) or a dedicated test PostgreSQL container via testcontainers.
• Mocks:
  • Mock OCR engines (pytesseract, pdfplumber) to avoid slow I/O during test suites.
  • Mock file uploads using fastapi.testclient.TestClient.
• Coverage Targets:
  • Engines: 100% logic coverage for TemplateRecognitionEngine and scoring weights.
  • Services: 90% coverage for CRUD operations.

Integration Testing

• Test end-to-end API flows:
  1. POST /api/documents/upload with a sample PDF.
  2. Wait for layout extraction.
  3. POST /api/templates to create a template.
  4. POST /api/templates/{id}/mappings/save to map extracted data.
  5. POST /api/documents/upload with a similar document to verify POST /api/documents/{id}/recognize returns the correct template match.

2. Frontend Testing (Angular)

Unit Testing

• Framework: Jasmine & Karma (or Jest if configured).
• Component Tests:
  • Verify TemplatesComponent renders the left and right panels.
  • Verify the AddField dialog toggles correctly and validates empty inputs.
• Service Tests:
  • Mock HttpClient using HttpTestingController to ensure TemplateService sends correct payloads to the backend.

E2E / Integration Testing

• Framework: Cypress or Playwright.
• Critical User Journeys (CUJ):
  1. User uploads a document, UI displays the layout preview visually.
  2. User creates a template and adds 3 fields.
  3. User drags a block from the Document Preview and drops it into a Template Field.
  4. User saves the mapping successfully.

3. OCR & Layout Extraction Accuracy Testing

Since the OCR engine relies on visual heuristics rather than AI models, testing its accuracy is crucial to prevent regressions.

• Golden Dataset: Create a dataset of 50-100 real-world business documents (Invoices, POs, Receipts).
• Evaluation Metric: Run the DocumentProcessor over the Golden Dataset and compare the output bounding boxes and classifications (HEADER, VENDOR, etc.) against manually annotated Ground Truth data.
• Acceptance Criteria: Maintain > 92% classification accuracy for logical block types.

4. Performance & Load Testing

• Tool: locust or k6.
• Scenario: Simulate 50 concurrent users uploading 2MB PDF documents simultaneously to ensure the DocumentProcessor does not exhaust server memory (OpenCV and Tesseract can be memory-intensive).
• Optimization: Ensure the process_file logic can be offloaded to Celery workers if the API starts blocking or timing out under load.