🎬 Por qué NO uso RAG (y qué uso en su lugar)
Video Details
Channel: Ruben Loan URL: https://www.youtube.com/watch?v=5xZxgqif-qg Relevance: ⭐⭐⭐
Summary
Contrarian perspective: Ruben Loan argues that for many use cases, RAG is overengineered. The alternative: context engineering (packing all relevant information into a large-context model’s context window rather than retrieving from an external store). Claims this is simpler, more reliable, and often produces better results.
PUMA Relevance
Important counterpoint for PUMA’s RAG justification. For PUMA Stage 1–3 (zero-shot, few-shot, CoT without RAG), the ‘context engineering over RAG’ argument applies: PUMA deliberately does NOT use RAG for Stage 1–3 to establish a clean baseline. The video’s argument also motivates testing whether Stage 4 RAG actually improves over Stage 3 CoT — that’s PUMA’s H2 hypothesis.