Agentic RAG Workshop

Why it matters

Agentic RAG moves the retrieval decision from the pipeline to the model. This workshop builds both halves, so you can see where the fixed retrieval chain ends and the agent’s choice begins.

What to practise

Chunk the knowledge base with RecursiveCharacterTextSplitter (size 800, overlap 120), embed it with NVIDIAEmbeddings into an in-memory FAISS store, and wrap the retriever in a ContextualCompressionRetriever with NVIDIARerank. Expose that chain through create_retriever_tool and hand it to create_react_agent. Serve the agent with langgraph dev, chat with it through the Simple Agents Client, and open its rag_agent trace in LangSmith. Then add a Tavily web_search tool over MCP and the list_available_skills and get_skill tools, and watch which tool the agent picks for each question, and in what order.

How it fits

The splitter settings are 6.1.1’s chunking decision, and the reranking chain is 6.1.3’s precision stage. The tool-choosing agent is 6.1.4’s retrieval as a tool; compare its open ReAct loop with 6.1.4’s planned graph and bounded correction.