RAG Is Over: RL Agents Are the New Retrieval Stack (inference.net)

Retrieval-augmented generation (RAG) — the embed-and-rerank stack (BM25/SPLADE, dense embeddings, RRF or cross-encoder re-rankers) — has hit diminishing returns, and agentic retrieval (LLMs given tools and run in loops) showed better results but was too slow and costly. The big shift: applying reinforcement learning to agentic agents makes them both superior and practical. Recent papers like DeepRetrieval (3B) and Search‑R1 show RL-trained small models can beat much larger frontier models and strong RAG baselines on benchmarks (DeepRetrieval reportedly outperforms GPT‑4o/Claude‑3.5 on 11/13 sets; Search‑R1 reports ~21–26% relative gains for 3B/7B over RAG across seven QA datasets), while caveats like benchmark contamination are noted. Technically, RL teaches small models to operate a toolbelt (grep, embedding search, SQL, web queries), decide when to search again, and stop when enough evidence is gathered. Reward shaping (penalize hallucinations and excessive tool calls; reward correct retrievals) produces models that are efficient, domain-aware, and adept at multi-turn query augmentation and evidence-seeking. The practical implication is an emerging architecture split: specialized, RL‑tuned retrievers for fast, domain-specific search feeding frontier models for final synthesis — lowering latency and cost while improving accuracy. Challenges remain: designing rewards, building evals, and deploying these systems effectively, but the trajectory suggests RL-powered agents will replace traditional RAG as the dominant retrieval stack.