Image OCR Extraction

# Image OCR Extraction

Extract machine-readable text from an image URL, preserve useful reading order, and report uncertainty without inventing content.

## Loadout capabilities

Use <integration-tag>direct:deepseek_ocr_tools</integration-tag> for OCR. The operational step is `direct:deepseek_ocr_tools:parse_image_to_text`, whose live schema requires one HTTPS `imageUrl`.

The Skill does not execute itself. Inspect the Action schema, then call it separately through Loadout.

## Supported scope

- Printed text in screenshots, scans, signs, receipts, and document images.
- A publicly reachable or otherwise authorized image URL.
- Plain machine-readable text returned by the live OCR Action.

Do not claim PDF parsing, local-file upload, handwriting accuracy, table reconstruction, formula extraction, bounding boxes, or layout visualization unless the live Action explicitly adds those capabilities.

## Workflow

1. Confirm the input is an image URL and that the user is authorized to send the image to the OCR provider.
2. Warn before processing credentials, identity documents, medical records, financial records, or other sensitive content. Ask for a redacted image when possible.
3. Inspect `direct:deepseek_ocr_tools:parse_image_to_text` and execute it with only `imageUrl`.
4. Check for an empty response, obvious truncation, or impossible character sequences.
5. Return the extracted text in a fenced block, followed by a short uncertainty note when needed.
6. If the user requests structured fields, derive them from the OCR output and clearly distinguish extracted text from inferred labels.

## Output format

```text
Source: [image URL]

Extracted text:
[verbatim OCR result]

Uncertainty:
[none, or concise notes about unreadable and ambiguous spans]
```

Preserve line breaks when they carry meaning. Never silently correct names, numbers, dates, amounts, or identifiers; show a proposed correction separately.

## Failure handling

- Invalid or unreachable URL: ask for a working image URL.
- Empty output: report that no readable text was detected.
- Low-confidence-looking output: identify ambiguous spans and ask for a clearer crop.
- Provider error: report the returned error and stop instead of fabricating text.

## Attribution

This is a native Loadout re-authoring of the OCR workflow from the pinned GLM-OCR source in [UPSTREAM.md](UPSTREAM.md). The preserved source license is in [LICENSE.txt](LICENSE.txt). The provider and supported feature set are intentionally different.