Does better OCR actually improve RAG retrieval?
Better OCR improves RAG only up to a point: once the characters are mostly right, retrieval quality is decided by reading order and page structure, and a high-accuracy engine that returns text in the wrong order still retrieves badly.
Character accuracy is the wrong headline number
OCR engines are compared on character or word error rate, which measures whether the letters are right. Retrieval does not care about individual letters; it cares whether a passage means what it appears to mean. An engine that reads every character correctly but returns a two-column page as interleaved lines produces passages that are locally perfect and globally wrong.
What actually breaks retrieval
Three things, in order of damage: reading order, because it decides which sentences end up in the same chunk; table structure, because a number separated from its heading is unrankable; and page furniture - headers, page numbers, watermarks - which repeats across every chunk and dilutes the vectors. None of these show up in a character error rate.
Blank pages are worse than failed pages
The dangerous outcome is an image that produces almost no text but is indexed anyway, because the resulting passages are nearly identical to one another. They then crowd real answers out of results. Building RAG Converter we found two unrelated images sharing about ninety percent of their chunk text, all of it boilerplate. The fix was to skip them and say so, rather than index a placeholder.
Resolution past the model's working size buys nothing
Detection models operate at a fixed working resolution and gain no accuracy from being handed more pixels. Measured on a page of rendered text, running at a native 1258 by 846 against the detector's own 736-pixel short side produced word-for-word identical output - the whole difference across the page was two full stops - and the smaller input was faster. A 12-megapixel phone photo is not fifteen times more readable; it is fifteen times more expensive.
Test on retrieval, not on transcription
The only benchmark that answers the question is your own: convert a sample, ask the questions you actually care about, and check whether the right passage comes back. An engine that wins on a transcription benchmark can lose on your corpus, and you will not find that out from anyone's accuracy table.