Agentic RAG

Plain RAG does one retrieval at the start; the model is a passive consumer of whatever the retriever returns. Agentic RAG lets the model decide when to retrieve, what to query, and whether it has enough evidence to answer — turning the retriever into a tool the model invokes mid-generation. This trades latency for accuracy on multi-hop queries where a single retrieval cannot surface all the needed evidence.

Adaptive-RAG

Adaptive-RAG (Jeong et al., NAACL 2024) addresses the simplest form of "when": route each query to one of three pipelines — no retrieval (closed-book), single-step retrieval, or multi-step retrieval — based on a learned classifier of question complexity. Easy questions don't pay the retrieval cost; hard ones get the full multi-hop budget. The classifier is trained on labels distilled from running each pipeline and observing which one was sufficient.

Auto-RAG

Auto-RAG (Yu et al., 2024) makes the retrieval loop fully driven by the LM. The model autonomously plans, retrieves, reflects on the snippets, and decides whether to issue another query — all expressed as natural-language reasoning. Training is via supervised fine-tuning on iterative RAG trajectories distilled from a strong teacher (GPT-4). The contribution over Self-RAG is that the loop is unconstrained — the model picks how many rounds it needs rather than emitting fixed reflection tokens.

Search-o1

Search-o1 (Li, Jin, Chen et al., 2025) is the agentic-RAG counterpart to o1-style reasoning models: long chain-of-thought interleaved with web searches. The model emits <|begin_search_query|> tags during its CoT; an external retriever runs the query and the system splices the results back into the trace before generation continues. A "Reason-in-Documents" module compresses the (often long) retrieved pages into the chain. The result: large gains on PhD-level science QA where one-shot retrieval is hopeless.

Search-R1

Search-R1 (Jin et al., 2025) replaces SFT/distillation with reinforcement learning to train the search-then-reason behaviour. The reward is just answer correctness on QA datasets — no per-step supervision. The model learns to invoke a search engine multiple times, even when the path is long and uncertain. This is the RLVR recipe applied to retrieval: the search calls become part of the policy, optimized end-to-end against a verifiable reward. Search-R1 outperforms RAG and Search-o1 baselines while being model-agnostic about the underlying retriever.

Reading list

• Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity — Jeong, Baek, Cho, Hwang, Park, NAACL 2024.
• Auto-RAG: Autonomous Retrieval-Augmented Generation for Large Language Models — Yu et al., 2024.
• Search-o1: Agentic Search-Enhanced Large Reasoning Models — Li, Jin, Chen et al., 2025.
• Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning — Jin, Yang, Su et al., 2025.

What to read next

• RAG — the static foundation this builds on.
• LLM Agents — the broader picture of tool-using LMs.
• RLVR — the RL machinery that powers Search-R1.