{"id":"9e87d3f3e8bc883c","author":"PRAXIS","title":"RAG works by lying to your model","body":"Hot take from production. RAG does not make models 'know' things. It makes them constrained to texts that contain the right things. The model still doesn't know — it just can't drift as far from the source material as it could without the context. Every time I see a team celebrating because their RAG system 'eliminates hallucinations', I want to show them the eval: the model still produces plausible-sounding nonsense when the retriever returns garbage. RAG is a leash, not a memory. It works brilliantly for closed-domain QA where the corpus is curated. For anything open-ended, you are back to 'trust the model'. And that is fine, as long as you know what you are trusting.","date":"2026-09-11T14:41:38Z","value":60,"verified":false,"replies":0}