13. Agentic RAG — Technical Knowledge Base¶

Overview¶

This section is a comprehensive 13-lesson technical deep-dive into building an advanced, production-grade Agentic RAG (Retrieval-Augmented Generation) system using LangGraph, LangChain, and ChromaDB.

The system synthesizes three research papers — Corrective RAG, Self-RAG, and Adaptive RAG — into a single, self-correcting pipeline that retrieves, grades, generates, and reflects on answers before returning them to the user.

Lesson Map¶

#	Lesson	Focus
01	Agentic RAG Architecture	System overview, research paper foundations, high-level flow
02	Improving RAG with Corrective Flow	CRAG paper — document grading and web search fallback
03	Boilerplate Setup	Poetry, dependencies, API keys, environment validation
04	Code Structure	Repository architecture — nodes, chains, tests, state
05	Vector Store Ingestion Pipeline	Document loading, chunking, embedding, ChromaDB storage
06	Managing Information Flow	GraphState definition and state management patterns
07	Retrieve Node	Vector store semantic search node
08	Relevance Filter for RAG	Retrieval grader chain, document filtering, testing strategies
09	Web Search Node	Tavily API integration, fallback retrieval
10	LLM Generation Node	Generation chain, RAG prompt, LCEL pipeline
11	Running the Complete Agent	Graph wiring, conditional edges, end-to-end execution
12	Self-RAG	Hallucination detection, answer validation, reflection loops
13	Adaptive RAG	Intelligent query routing, conditional entry points

Architecture at a Glance¶

flowchart TD
    START(("▶ START")) --> ROUTE{"🔀 Route Query<br/>(Adaptive RAG)"}
    ROUTE -->|"vectorstore"| RETRIEVE["📥 Retrieve"]
    ROUTE -->|"websearch"| WS["🌐 Web Search"]
    RETRIEVE --> GRADE["📝 Grade Documents<br/>(Corrective RAG)"]
    GRADE -->|"all relevant"| GEN["🤖 Generate"]
    GRADE -->|"some irrelevant"| WS
    WS --> GEN
    GEN --> REFLECT{"🔍 Reflect<br/>(Self-RAG)"}
    REFLECT -->|"hallucinated"| GEN
    REFLECT -->|"useful"| END(("⏹ END"))
    REFLECT -->|"not useful"| WS

Usage¶

Start with Lesson 01 for the conceptual overview, then follow sequentially. Each lesson builds on the previous one, incrementally constructing the full system.