ray🖤🇰🇷

Your "RAG is Dead" take is wrong; but keep thinking it because it gives me job security!  🔥 Every time a new LLM drops with a bigger context window, tech influencers rush to declare "RAG is dead!" without understanding the reality of production AI systems. After building ML systems at Google and LinkedIn, let me set the record straight: - Fiction.liveBench data shows even Llama 4 (with its claimed 10M token context) fails at basic comprehension beyond a few thousand tokens. - Most enterprises have terabytes of documents. No context window can encompass a pharmaceutical company's 50K+ research papers and decades of regulatory submissions. - The economics are brutal: Glean's research shows each token adds ~0.24ms to response time. A 100K token context means 24 seconds of waiting! At $0.01/1K tokens (GPT-4), that's $1 per query just to process input. - Engineering reality: Separating retrieval from generation creates maintainable systems you can actually debug when things go wrong (and they will). Certainly, fewer use cases need RAG now; we have definitely exited the era where simple QA on a single PDF requires RAG. The "RAG is dead" narrative fundamentally misunderstands AI system design. When you build systems at scale, you quickly learn that retrieval is an essential tool for many systems.