How much of your own data needs to be part of the response?

This shapes whether retrieval or fine-tuning is needed.

What are your latency requirements?

Response time shapes which architectures are practical.

How much does accuracy matter relative to cost and speed?

There are real trade-offs here — no approach gives you all three at once.

Does the task require multiple reasoning steps or external tool calls?

This is the clearest signal for whether you need a simple prompt or a full agent architecture.