RAG

RAG answers questions from documents the model was never trained on: retrieve relevant chunks, stuff them into the prompt, generate an answer grounded in them. Almost every AISL course and workshop touches it. The material's central claim about RAG is also its most useful lesson: a fixed pipeline searches once and hopes, and the interesting failures begin exactly there.

The classic pipeline

The recipe appears first in AI Hero: 7-Day AI Agents Crash-Course: download a real GitHub repository, chunk the documentation, index it, retrieve, generate with citations. The AI Engineering Buildcamp module 2 repeats it over real use cases — a FAQ chatbot, content management, YouTube transcripts, PDF books — with a reusable RAG class.

Agentic RAG builds the classic version first, over the Evidently AI documentation with minsearch, precisely so you can watch it break.

Where fixed RAG breaks

Two failures show up consistently across the material:

• A typo or unusual phrasing derails keyword retrieval, and the pipeline has no second chance. The Agentic Loop from Scratch starts from exactly this failure.
• Chunking loses context: the retrieved snippet is not the document, and the answer inherits the gap.

Retrieval quality itself is a separate discipline — see Vector Search for TF-IDF, boosting, and embeddings.

Agentic search: RAG with tools

The fix the material converges on is to stop fixing the pipeline and give the decision to the model. In Agentic RAG the agent gets search (highlighted snippets) and get_file (the full document), and mirrors how a human reads docs: search, scan, open the promising one. The same pattern powers the FAQ agents in End-to-End Agent Deployment and Building AI Agents with MCP, PydanticAI, and OpenAI.

The buildcamp module 3 formalizes it as agentic RAG with tool calling, on the way from RAG into full Agents.

Where to go next

• The retrieval layer underneath: Vector Search
• Giving the model control of the loop: Agents
• Keeping a RAG agent on topic and safe: Guardrails