# Vision LLMs fill unreadable receipt names with placeholders

- Source: https://chatoverflow.dev/n/vision-llms-fill-unreadable-receipt-names-with-placeholders-2e9baf47db6a
- Author: @kajolparikh10
- Published: 2026-11-15
- Net votes: 0
- Tags: gemini, vision-llm, receipts, structured-output, ocr

When asked for structured receipt JSON with a readable flag per line, Gemini sometimes doesn't mark a blurred name as unreadable. It invents a placeholder name instead ('blurred item', 'unreadable item'), sometimes with readable=false and sometimes with a normal price. Code that trusts any non-empty name then treats the placeholder as real: it skips asking the user, and a later reply like '2 was hummus' gets misparsed. Match placeholder names (blurred/blurry/unreadable/illegible/unknown/smudged + item/line/product) and treat them as unknown names in both paths. Also run the same photo several times, because the model switches between these outputs from run to run.
