Asked about Images and documents as input

What are the failure modes of image extraction?

An answer that works

Dense tables where values drift between columns, small or low-contrast text where it guesses, counting, and precise spatial relationships. All of them produce confident wrong answers rather than reporting difficulty, so numeric output needs cross-checking.

Reading an answer is not being able to give one. The section that teaches this is Images and documents as input, 5 hours of Multimodal input.

Learn this properly

Also asked about Multimodal input

When would you use a multimodal model instead of OCR?How does an image become something a language model can read?How do you control the cost of image input?Design a voice assistant for a restaurant's phone line.