| name | text-and-ocr-analyzer |
| type | tool |
| description | Extracts and analyzes text within images to verify spelling, placement, and clarity. |
| system_prompt | You are an expert OCR and text analysis AI. Extract all visible text from each image provided. For each piece of text, return:
- the exact characters (preserve case and punctuation)
- whether the spelling is correct for natural English (or the language of the prompt)
- approximate placement on the image (top/middle/bottom × left/center/right)
- any obvious rendering issues (blurry, partially occluded, distorted)
Be conservative: if you cannot read a character with high confidence, mark it as `?`.
|
| input_schema | {"images":"list[base64_str]","query":"str"} |
| output_schema | {"image_1_text":"str","image_1_spelling_correct":"bool","image_1_placement":"str","image_2_text":"str","image_2_spelling_correct":"bool","image_2_placement":"str"} |
text-and-ocr-analyzer
When to use
Whenever the prompt asks to add, modify, or remove text on an object (signs, labels, captions, t-shirts, posters, license plates, etc.). Many editing models hallucinate plausible-looking but misspelled text — holistic visual evaluation routinely misses this.
Example query
What text is on the storefront sign in each image? Is it spelled exactly as "Sale"?
Reading the output
Treat a False in image_*_spelling_correct as strong evidence of instruction-following failure (drop the candidate's instruction-following score by at least 1 point). Treat mismatched placement from the prompt's request as a separate, smaller deduction.
Cost
This tool runs one extra VLM call per evaluation, so don't invoke it unless the prompt explicitly contains text-editing intent.