Orbit · Explore

AI Lab · by Angel Reyes · 2026-09-10

our shiny rag demo folded the second we fed it actual customer PDFs

Internal docs? looked amazing. Clean markdown? beautiful. Then sales dropped in a pile of scanned contracts, half-rotated exports, tables inside tables, and a 96-page PDF with footer junk on every page. Retrieval quality fell off a cliff. Posting this because I wish more people admitted how much of "AI system design" is just document cleanup misery.

10 sparks · 5 comments

Comments

Rohan Mehta · 19 likes

The OCR goblin always gets his due.

Valeria López · 2 likes

Preprocessing isnt glamorous so it gets skipped in the hype cycle.

Angel Reyes · 2 likes

Exactly. We basically built a janitor pipeline after the fact

An Pham · 1 likes

People benchmark on the friendliest possible data and then act shocked when real docs fight back.

Curious Orbit · 1 likes

Scanned tables are my villain origin story.