Shotnest · Library
Understand a short classification result
Text under 12 characters after trimming goes straight to Other with score 0: “Receipt 123” has 11 characters, “Receipt 1234” scores 2.
The 12-character cutoff
Before any keyword is checked, the classifier lowercases the text, trims spaces at both ends and counts what is left. Under 12 characters, it returns Other with a score of 0 and does not look further. “Receipt” has 7 characters, so it ends there even though receipt is a Receipts keyword worth 2 points in longer text.
The same limit appears in the app. Shotnest only asks Apple’s on-device model for a second opinion when the text is at least 12 characters long and the rule score is under 3, so very short text stays on Other.
Try the original example
- Text
- Receipt
- Expected shelf
- Other
Test the boundary
- Enter “Receipt”: 7 characters, Other, score 0.
- Enter “Receipt 123”: 11 characters, still Other, score 0.
- Enter “Receipt 1234”: 12 characters, the keyword now counts and Receipts scores 2.
- Note that 2 is below the threshold of 3, so the suggestion is not confident.
- “Receipt” · 7 characters
- Other · 0
- “Receipt 123” · 11 characters
- Other · 0
- “Receipt 1234” · 12 characters
- Receipts · 2
The cutoff is about characters, not about how strong the words are. “Total 14.25” also has 11 characters and returns Other with score 0, even though it contains a Receipts keyword and a number that looks like an amount.
What to do with a short result
Add the real surrounding text from the screenshot, such as the merchant line or the total, instead of typing extra words to reach the limit. If the screenshot genuinely contains almost no text, Other is the honest answer and the image itself has to be looked at.
- Counting the characters of the untrimmed text, with leading spaces.
- Padding a short text with invented keywords to force a shelf.
- Reading Other with score 0 as a judgement on the screenshot’s value.
Because the check uses trimmed length, spaces or empty lines around the text never help it reach the limit.
Limited evidence, stated plainly
Score 0 means the rules had too little text to work with. The classifier tool linked below reports that fallback; it does not look at the image to find text the copy missed.