Testing
Strategy
The OCR pipeline is tested at multiple levels:
- Unit tests — individual stage logic (quality detection, classification rules, chunking strategies)
- Integration tests — end-to-end pipeline runs on curated test documents
- Quality benchmarks — OCR accuracy against ground truth, retrieval precision/recall
- Manual review — human inspection of RAG answers with citations
Test Documents
A representative set of archive documents covering all target document types:
| Document Type | Count | Characteristics |
|---|---|---|
| Printed book page | 5 | Clean, consistent font |
| Typewritten letter | 5 | Variable quality |
| Handwritten record | 5 | Cursive, varied hands |
| Ledger page | 3 | Tables, numbers |
| Invoice/Check | 3 | Structured, amounts |
| Meeting minutes | 3 | Paragraphs, dates |
| Form | 2 | Field groups |
| Poor-quality scan | 3 | Blur, skew, damage |
Quality Metrics
- OCR character accuracy: CER against ground truth transcription
- Classification accuracy: % correct document type assignment
- Metadata precision/recall: dates, people, organizations
- Retrieval precision@k: % of top-k chunks relevant to query
- Answer groundedness: % of RAG claims supported by source citations
See test-plan.md for strategy and test-cases.md for concrete test vectors.