Module 4 · Project 3: Agentic RAG Helpdesk

OBJECTIVE: upgrade the policy Q&A assistant from the LLM module into an agent that decides when and where to search (policy documents, the shop database, the web), checks whether what it found is relevant, searches again if needed, and checks its own answer — and learn agentic RAG. WHAT WE COVER (in project order): • Why plain RAG fails on harder questions • Agentic RAG: retrieval as a tool the agent chooses • Routing between sources: vector store, SQL/MCP tools, web search • Corrective loop: grade documents → rewrite the query → search again • Self-check of the final answer (grounded or not) • Single-agent vs multi-agent agentic RAG • Glimpse: GraphRAG and knowledge graphs • Evaluation against plain RAG, and deployment AFTER THIS MODULE YOU CAN: • Build an agent that searches several sources and corrects itself • Show with numbers when agentic RAG is worth its extra cost DATA: the policy PDFs and 50-question test set from the LLM module + the shop database (via your MCP server) + 20 new questions that need two sources or a second search. TOOLS: LangGraph, ChromaDB, your MCP server, a web-search API.