Opening the source...
OCR extracts text from page images using AI vision. Pages are processed in batches of 5, with multiple batches running in parallel for speed.
Batch sizes 1, 5, and 10 show no significant quality difference. Quality only degrades at batch size 20 (30-1 loss vs batch 10, p < 0.0001).
We ran an OCR quality experiment on 31 pages from a Latin manuscript, testing 8 conditions:
An AI judge compared outputs pairwise across all conditions. Results were analyzed using ELO ratings and statistical significance testing.
| Comparison | Winner | Score | Significant? |
|---|---|---|---|
| Batch 1 vs 5 | Batch 1 | 17-14 | |
| Batch 5 vs 10 | Batch 5 | 18-13 | |
| Batch 10 vs 20 | Batch 10 | 30-1 | |
| Simple vs Elaborate prompt | Tie | 17-14 |
Recommendation: Use 10 parallel batches for maximum speed. Quality is statistically equivalent to sequential processing (batch 1). Only batch 20 shows degradation.
Translation converts OCR text to English. Unlike OCR, translation benefits from context because it needs to maintain:
Pages are processed sequentially in batches of 5. After each batch, the last page's translation is passed as context to the next batch, creating continuity.
OCR is image-based. Each page is processed independently using visual recognition. Context from previous pages doesn't significantly improve accuracy.
Translation is text-based. The AI needs context from previous pages to maintain consistency in terminology, style, and handling of cross-page sentences.
Experiment conducted on a Latin manuscript using AI-assisted judging.