I Went Looking for the Dead Internet in My RAG. I Found a Bad Experiment.

I removed duplicate sources from a RAG benchmark and watched held-out accuracy fall from 10 of 20 tasks to 5 of 20. The tempting conclusion was that source deduplication made retrieval worse. The audit found another change hiding in the comparison. Lexical top-k supplied roughly five documents; the…

Read the full story at DEV Community — Machine Learning ↗

Timeline · 1 report

1. 2026-11-03 08:36 · DEV Community — Machine Learning
   I Went Looking for the Dead Internet in My RAG. I Found a Bad Experiment.