10 decisions
The choices that shape the build
Every page here gives a verdict. “It depends” is what you already knew — the useful part is what it depends on, and then an answer for the common cases.
- AgentvsFixed workflowUse a fixed workflow unless the sequence of steps genuinely cannot be known in advance.Most asked
- Hosted APIvsSelf-hostingUse an API unless you have high, steady utilisation or a hard data-residency requirement.Most asked
- RAGvsFine-tuningUse retrieval when the model needs facts it does not have, and fine-tuning when it needs to behave differently.Most asked
- Few-shotvsZero-shotStart zero-shot and add examples only when the failure is format or ambiguity.
- Large modelvsSmall modelDo not choose — route.
- Long contextvsRAGRetrieve rather than stuff, in almost every case.
- Prompt engineeringvsFine-tuningExhaust prompting first — it is free, immediate, and solves most problems.
- Reasoning modelvsStandard modelUse a reasoning model for genuinely hard problems and a standard model for everything else.
- Semantic searchvsKeyword searchUse both.
- StreamingvsBatchThese are not competing options, they are different traffic.
Why these have verdicts
Most comparison articles in this field end without answering the question, which is safe for the writer and useless for the reader. Every page here commits: use retrieval unless the problem is behaviour; use a workflow unless the sequence is genuinely unknowable; route between model tiers rather than picking one.
Where the answer genuinely turns on your numbers — which is often — the page says which numbers, and links to whichever of the 102 tools produces them. A decision you can compute beats a decision you have to feel.
For the decision between specific models rather than approaches, the price table and the model finder are the faster route.