Someone told you to build RAG. Should you?
It’s the default answer now. Documents plus a language model equals retrieval-augmented generation, and the architecture diagram draws itself.
Sometimes that’s right. Often the corpus is small enough to put in the prompt, or the real requirement is a database query rather than a search, or the honest problem is that nobody has decided who maintains the index in a year.
This site is about the decision — not how to build it, whether to.
What you’ll find here
- When not to use RAG. Four situations where retrieval is the wrong shape for the problem, and what to build instead.
- RAG versus long context. Both work. They have completely different cost curves, and the crossover depends on three variables you can actually estimate.
- RAG versus fine-tuning, as an ownership question. Not “which is more accurate” — which one leaves you with a system you can update on a Tuesday afternoon.
- What it costs to run. Embedding, storage, inference, reindexing, and the two line items nobody budgets: evaluation and the person who owns the corpus.
- Latency budgets. Where the milliseconds go, and which stages you can cut when your target is under two seconds.
The method throughout: name the criteria, work the arithmetic, then commit to a recommendation. A survey of options is not useful to someone who has to choose one.
Written for the person accountable for the budget and for the thing still working in eighteen months.
No vendor prices quoted as fact — they change monthly and vary by contract. You get the cost model and clearly labelled example numbers; you substitute your own and the arithmetic still works.
Latest posts
-
When Not to Use RAG
Four situations where retrieval is the wrong shape for the problem, what to build instead, and the threshold at which each answer flips.
-
RAG Versus Long Context: The Cost Arithmetic
Both architectures work. They have different cost curves, and the crossover depends on three variables you can estimate this afternoon.
-
RAG Versus Fine-Tuning as an Architecture Decision
Not which is more accurate — which one leaves you able to fix a wrong answer on a Tuesday afternoon. An ownership-first comparison.
-
What a RAG System Actually Costs to Run
The full line-item breakdown: embedding, storage, inference, reindexing — plus the two costs nobody budgets and one that dominates everything.
-
Latency Budgets for Retrieval
Where the milliseconds go in a RAG request, which stages you can cut when your target is under two seconds, and what each cut costs you.