RAG for images and scans: why OCR is the whole job
An image stores its words as pixels rather than text, so it yields nothing for RAG without OCR - and a converter that indexes it anyway makes the corpus worse, not better.
Decoding an image is not reading it
A converter can open a JPEG, confirm it is valid, measure it and hash it, and still have no idea what it says. Decoding gets you pixels. Reading needs optical character recognition, which is a separate model doing a genuinely different job.
Why indexing an unreadable image is worse than skipping it
Without OCR the only text a converter can attach to an image is boilerplate - the filename and an empty template. Do that across a folder and every image produces nearly the same passage, so they all match every query equally and push real answers down the results. Measured on two unrelated images, roughly 90% of the generated text was identical.
What to do instead
Run OCR before converting, or use a converter that does it server-side where the model can run at full speed. RAG Converter's browser tier skips images and lists them as skipped, so you know they are missing rather than discovering it through bad answers weeks later.
OCR quality varies enormously by source
A clean 300 DPI scan of printed text is close to solved. A phone photograph at an angle, a handwritten note, or a screenshot of a dark-mode terminal are all substantially harder, and the resulting text quality caps how good retrieval over it can ever be.