Image OCR by langbot-team

# Image OCR

Extract text from images and scans with Tesseract OCR.

## What it does

- Reads text from photos, screenshots, and scanned documents.
- Optional preprocessing (grayscale, threshold, deskew) for noisy images.
- Multi-language, including English and Chinese (eng+chi_sim).
- Reports low-confidence regions instead of guessing.

## Files

| File | Purpose |
|------|---------|
| `SKILL.md` | Instructions the agent receives on activation |
| `scripts/ocr.py` | Preprocess + OCR an image with language/confidence options |
| `references/accuracy.md` | Preprocessing and language-pack tips |

## Requirements

Installs `pytesseract` + `Pillow`; needs the `tesseract-ocr` binary and the
relevant language packs (e.g. `tesseract-ocr-chi-sim`) available in the sandbox.

## License

Apache-2.0.


---
name: image-ocr
display_name: Image OCR
description: "Extract text from images and scanned documents using OCR. Use when the user sends a photo, screenshot, or scan and wants the text read out, transcribed, or extracted — including receipts, business cards, whiteboards, slides, or any picture containing text. Supports multiple languages including English and Chinese. Do NOT use for editable PDFs with a real text layer (extract directly instead)."
license: Apache-2.0
---

# Image OCR

Read text out of an image or scanned page using Tesseract OCR in the sandbox.

## When to use

The user provides an image (photo, screenshot, scan) and wants the text it
contains — transcription, extraction, or "what does this say?".

## Execution steps

1. **Locate the image** under `/workspace`. Confirm the language(s) of the text;
   default to English + Simplified Chinese (`eng+chi_sim`) when unsure.
2. **Preprocess** for better accuracy with `python scripts/ocr.py <image> --preprocess`.
   It grayscales, thresholds, and deskews using Pillow before OCR (see
   `references/accuracy.md` for when each step helps).
3. **Run OCR**: `python scripts/ocr.py <image> --lang eng+chi_sim`. The script
   installs `pytesseract` + `Pillow` and ensures the Tesseract binary and the
   requested language packs are present.
4. **Return** the extracted text. For structured docs (receipts, tables), preserve
   line breaks and note low-confidence regions.

## Rules

- Pick the right `--lang`; the wrong language pack wrecks accuracy.
- Preprocess noisy/low-contrast images before OCR; skip it for clean screenshots.
- Never fabricate text you cannot actually read — mark unclear spans as `[?]`.

## Available resources

- `scripts/ocr.py` — preprocess + OCR an image, with language and confidence options.
- `references/accuracy.md` — preprocessing and language-pack tips for better results.