What's the smallest RAG pipeline that actually works — and why is it exactly one LLM call?

A base model knows nothing about your private documents, your internal catalog, or anything that postdates its training cutoff. Ask it anyway and you get one of three failures: it hallucinates a plausible answer, it leans on stale parametric memory, or you try to stuff the whole corpus into the prompt and blow the context window.

The fix is not a bigger model or a bigger prompt. It is a disciplined four-stage flow with exactly one LLM call.

Query → Retrieve → Augment → Generate

Query. Take the user's question as-is.

Retrieve. Embed the query and pull the top-k most similar chunks from your vector database. This is the only place new knowledge enters the system.

Augment. This is the step everyone underestimates. The retrieved chunks must be injected into a grounding prompt template — a structured prompt that tells the model to answer only from the provided context and to say so when the context is insufficient. Appending the chunks to the end of a freeform prompt is not augmentation; it is hoping.

Generate. One LLM call produces the answer from the grounded prompt. One call — not a chain, not an agent loop.

Why the single-call invariant matters

The single-LLM-call invariant is what makes traditional RAG cheap, fast, and debuggable. Every answer traces to one prompt you can inspect: the template plus the retrieved context. When an answer is wrong, you look at exactly two things — did retrieval surface the right chunks, and did the template force the model to use them? Multi-hop and agentic pipelines have their place, but most "we need RAG over our docs" problems are solved by getting this one-call flow right first.

If answers still drift after this, the problem is almost never the model. It is that retrieval is not being injected into a real grounding template — so fix the augmentation step before you reach for anything fancier.