Naive RAG — embed, retrieve, stuff, generate — was version one of a design that has since grown through three generations: advanced RAG with corrections around the retrieval step, modular RAG with swappable stages, and agentic RAG where the model decides when and what to retrieve at all.

The generations are responses to observed failure, not fashion. Naive pipelines retrieved once and trusted blindly, so wrong retrievals became confident answers. Advanced RAG added pre-retrieval improvements (query rewriting, expansion) and post-retrieval corrections (reranking, compression, relevance filtering). Modular RAG — the Gao et al. (2023) survey formalized it — reframed the pipeline as composable modules: retrieval, memory, routing, fusion, scheduling, each augmentable or rearranged. Self-RAG trains the model to critique its own retrieved evidence; corrective RAG (CRAG) runs a lightweight relevance evaluator that triggers web search or re-retrieval when the vector store comes back weak.

Agentic RAG completes the arc: retrieval becomes a tool the model calls strategically — iterating, refining queries, reading slowly where it matters — instead of a fixed stage that always runs. The cost is unpredictability and spend; the benefit is that hard multi-hop questions, which defeat any single-shot pipeline, finally get answered by decomposition and persistence. Production systems increasingly run both: a fixed fast pipeline for the 90%, an agentic path escalated for the rest.

Choosing a Generation

• Naive: corpora that are small, curated, and phrased like the questions asked of them. Ship it, measure it, expect to outgrow it.
• Advanced: the default once query logs show vocabulary mismatch, stale chunks in answers, or rerank-worthy near-misses at the top.
• Agentic: multi-hop questions, heterogeneous sources, or freshness requirements — accept variable latency and cost, and cap it with budgets.

The through-line is a single idea: retrieval stopped being plumbing and became reasoning. Every generation moved the "when to look up what" decision closer to the model — and every generation raised the value of the evaluation harness that tells you whether the change helped.