Same model, same prompt, worse API extraction: apps do invisible pre-processing you're skipping

TangeloOk9486 · reddit · 2026-09-23

A Reddit post explains why document extraction via raw API often underperforms the ChatGPT/Gemini apps with identical models and prompts: apps silently downsample PDF pages (768px short side / 90dpi for ChatGPT vs 210dpi for Gemini, plus mediaresolution controls), pull embedded text layers with table structure intact, use parsers like docling or llamaparse, and chunk+retrieve instead of stuffing whole PDFs into context. To match app quality on the API, replicate the pre-processing yourself: render higher dpi where vision matters, extract the text layer, feed clean markdown, and rebuild the scaffolding.

Original post →

More from coding & agent

coding & agent channel →