# OCR

The primary artifact is `ocr.jsonl` còn `artifact_metadata.jsonl` records the input, model, prompt version, subset, and Git lineage

## `ocr.jsonl`

Each line is one `OCRFragment`, a visible text region in one keyframe.

Common fields:

- `fragment_id`: OCR observation identifier
- `video_id`: source video identifier
- `time_span`: `[timestamp_ms, timestamp_ms]` in source video time
- `text`: recognized visible text
- `bbox`: normalized text region with `x_min`, `y_min`, `x_max`, and `y_max`

There is no placeholder for frames without text and no fabricated confidence score. Join OCR with captions or detections on `video_id` and timestamp; compare normalized bounding boxes when spatial context is needed.

## Generation flow

HunyuanOCR reads the canonical keyframe images, emits visible text regions, and the job validates each result as an `OCRFragment`. It publishes one flat JSONL file and removes provider work files and the retired bundled OCR representation after a successful run.

Run OCR for the original keyframe set:

```bash
make data-ocr INPUT=/data/aic/shared/artifacts/keyframes/keyframes.jsonl OUTPUT=/data/aic/shared/artifacts/ocr SUBSET=700vid
```