A data scientist ran the blind test everyone's been arguing about: RAG versus Kimi K3's 1M-token context window on the same 12 questions with the same model. Long context won decisively on quality—perfect scores on all 12 questions versus RAG's near-perfect but noticeably weaker completeness. The catch: it cost 16 times as much and took 3 times longer per question. The test used a 127,068-token corpus (32 articles) that occupied only 12% of the available window, which matters because that's the safe zone with real evidence behind it. Beyond roughly a fifth of the window, you're extrapolating into untested territory where the model has to get selective again. The roadmap boils down to matching the tool to the problem shape. For small, rarely-queried corpora that fit comfortably in the context window, skip RAG entirely—the long-context path is simpler and cheaper. For frequently-queried knowledge bases, RAG still wins on total cost despite lower per-query quality. The real economics hinge on query volume, not corpus size: at 12 questions the cost difference is negligible, but scale to 12,000 queries and it becomes the entire decision. Watch for three failure modes: reasoning tokens silently eating your answer budget, prefix caching that only works roughly a third of the time, and confident wrong answers when the system should admit gaps. When tasks require pulling from dozens of unconnected sources—competitor research, market scans, documents nobody's assembled—neither RAG nor long context cuts it. That's where Kimi K3's Agent Swarm handles the orchestration: decomposing the task, spinning up parallel sub-agents, running coordinated steps, then stitching separate findings into one connected structure instead of handing back 300 isolated boxes. The mistake is picking one tool and forcing everything through it. Real projects use all three: RAG for frequent, narrow lookups on large stable corpora; long context for occasional deep reads of one corpus; Agent Swarm for wide research where connections matter.