What Is Adaptive RAG

Where we left off

If you've completed naive_rag, hybrid_rag, graph_rag, corrective_rag, and agentic_rag, you've built five different ways of turning a question into a grounded answer, and watched each one fail in its own specific way. Every one of those courses still made a choice before you ever asked it a question: which retrieval strategy to run, fixed, for every question that strategy's lesson threw at it. That was the right call for teaching each strategy in isolation. It stops being the right call the moment two different questions, asked back to back, need two different strategies to answer well.

Adaptive RAG is the idea that a system should look at a question first, decide what kind of question it is, and only then pick how to retrieve for it. Not every question needs the same retrieval strategy: a one-fact lookup and a question whose answer is scattered across two documents are not the same problem, and running both through the same fixed pipeline treats them as if they were.

A multi-hop question, against a fixed single-document strategy

This lesson reuses naive_rag's own Lesson 8 pipeline, chunk, embed, retrieve top-1, generate, unchanged, and asks it a question this course's fixtures were built to expose: "What two hobbies happen in the same room as the weather station?"

The answer needs two files at once. bookshelf.md says the study has a bookshelf (one hobby: organizing it) and that the weather station's Raspberry Pi lives there too. cello-practice.md independently confirms the same room and names the other hobby: cello practice. Neither file alone names both hobbies, only together do they answer the question fully.

Top-1 retrieval, by definition, hands the model exactly one of those two files. Whichever one scores higher this run, the model can only answer with what that single file says, or, just as likely, conclude the context doesn't mention any hobbies at all, because the file it got doesn't actually say "hobby" anywhere in those words. Either outcome is the same underlying problem: a fixed k=1 naive strategy structurally cannot see two documents at once, no matter how the prompt is worded.

The code, piece by piece

This lesson's lesson.py is naive_rag Lesson 8's retrieve() and generate_answer(), byte-for-byte the same shape, run once against a question chosen specifically because that shape can't answer it well. Nothing here is new mechanically; what's new is diagnosing why it comes up short, and naming that as a routing problem rather than a retrieval-quality problem. The chunking, embedding, and generation code already works correctly, individually. The failure is entirely in k=1 being the wrong shape for this question.