Not every question deserves the expensive machinery
Agentic RAG (Lesson 17) can answer questions naive RAG cannot, but at real cost in latency and compute. Running every question, including "what is your return window," through a multi-step agentic loop wastes both, for no benefit, since a single retrieval already answers it correctly. Adaptive RAG exists to avoid exactly that waste.
How it works
Before any retrieval happens, a lightweight classifier (often a small, fast model, not the main generation model) assesses the incoming question's complexity: does it look answerable directly from the model's own knowledge with no retrieval at all, does it look like a straightforward single-document lookup, or does it look like it needs multi-step, agentic reasoning? The question is then routed to the cheapest architecture actually capable of answering it correctly: no retrieval, naive/advanced RAG, or full agentic RAG, respectively.
How this differs from modular RAG's routing
Modular RAG (Lesson 8) routes by data source, a database question goes to SQL, a document question goes to vector search. Adaptive RAG routes by complexity, a simple question gets a cheap, fast path and a complex one gets an expensive, thorough path, independent of which data source either one touches. The two routing dimensions are complementary and commonly combined: a production system can adaptively choose "how much retrieval effort" while modularly choosing "which retrieval tool," for the same question.
Checkpoint
- Adaptive RAG classifies query complexity before retrieval and routes each question to the cheapest architecture capable of answering it correctly.
- This avoids running simple questions through expensive agentic machinery, and avoids running complex questions through a single-shot retrieval that cannot handle them.
- Adaptive routing (by complexity) and modular routing (by data source) are different dimensions and are commonly combined in the same system.
If anything here still feels unclear, ask before moving to Lesson 21.