RAG vs Long Context vs Fine-Tuning: Which One Do You Actually Need

All three solve the same problem — get domain-specific or up-to-date information into a model's answers — but they trade off differently: Long context (just paste it in): cheapest to build, works immediately, and current-generation context windows are large enough for most single-session tasks. Best when the relevant material fits in one request and doesn't need to be searched across a much larger corpus. RAG (retrieve, then paste in): best when your knowledge base is larger than any context window, changes often, and you need to cite which document an answer came from. Cost is mostly engineering (a retrieval pipeline, an index to keep in sync) rather than model cost, and quality is capped by retrieval quality — a great model with bad retrieval still gives bad answers. Fine-tuning: best when you need to change how the model behaves (tone, format, a narrow skill) rather than what it knows — fine-tuning is a poor way to teach new facts (they can leak, go stale, and are hard to update) but a good way to teach a consistent style or task pattern. It's also the slowest and most expensive to iterate on. Default order to try, cheapest first: long context, then RAG, then fine-tuning. Skip straight to RAG if your corpus won't fit in context; skip straight to fine-tuning only if the problem is genuinely about behavior, not knowledge.