Spot the preprocessing mistakes that doom this receipt-reading pipeline

Receipts are the hardest input for a vision model because everything important is tiny: prices, item codes, totals. The pipeline must protect those small pixels. Line two forces the receipt into a 512 by 512 square. Real receipts are tall and narrow. Squashing them warps every character and shrinks the text below what the encoder can resolve. The fix is an aspect-preserving resize, with the long edge near the provider maximum (around 1500 px). Line three saves at JPEG quality 20. That smears the edges of small digits, which is exactly the information you need. And it saves nothing on cost: vision providers bill by pixel dimensions, not by bytes. Use quality 90 plus, or PNG for OCR-heavy work. Line four asks for low detail mode. Low detail gives the model a fixed small token budget, intended for a quick scene description, not for reading dense text. High detail is the right call for extraction. Fix the three settings together and the same VLM goes from hallucinated totals to clean line items.

**Three independent bugs**, each attacking small-text legibility, each easy to mistake for a sensible default. The **square resize** to 512 by 512 destroys aspect ratio. Receipts are tall; you have just crushed 8x the height into 1x the width, and every glyph along with it. Use an aspect-preserving resize where the long edge sits near the provider maximum, often around 1500 px for high-detail paths. That keeps line items legible after preprocessing. **JPEG quality 20** is a category error. The intuition is that smaller bytes mean cheaper calls, but vision token cost follows pixel dimensions, not file size. The compression buys you nothing on cost and aggressively smears character edges. Use quality 90 plus, or lossless PNG when the document is OCR-heavy. **Low detail mode** is the third trap. Providers expose a low detail path that maps the image to a small fixed token count, intended for scene gist. Reading dense receipts needs the high detail path, which sends a much larger token budget so small text can actually be resolved. Fix the three together and you stop fighting the model. The same VLM that hallucinated totals will quietly read every line.