Agentic RAG
Agentic RAG: a planner agent decides actions, calls tools (retrieve, web search, DB, calculator), collects and reasons, and loops until it can answer, with the full architecture, capabilities, a 9-step walkthrough, a worked example, the detailed agent loop, components, benefits, challenges and best
Cheat sheet · Offered as a template
In this diagram
- Agentic RAG
- What is Agentic RAG?
- Uses an agent (LLM) to reason, plan and take actions.
- Dynamically decides what information to retrieve, from where and how.
- Iteratively retrieves, reasons and generates better answers.
- Best for complex, multi-hop and open-ended questions.
- Overview (Flow)
- User Query
- Agent (Planner)
- Decide Action
- Retrieve
- Search Web
- Query DB
- Calculator
- … (Other Tools)
- Collect & Reason
- LLM (Generator)
- Answer
- Architecture (Flow Diagram)
- Understand intent, entities, constraints and context
- Query Understanding
- Plan steps, choose actions and tools to use
- Tools / Actions
- Retriever (Vector DB)
- Keyword Search (BM25)
- Web Search
- SQL / DB Query
- Calculator Code Exec
- … Others
- Observation (Results)
- Get results from tools / data sources
- Agent (Reasoning)
- Reason over results, decide next steps or finalize
- Memory (Conversation + Intermediate)
- Store context, history and intermediate findings
- Generate final answer using gathered context
- Answer (Response)
- Return final response to the user
- Key Capabilities
- Autonomous decision making (what to retrieve, from where, how many times)
- Multi-step reasoning and iterative retrieval
- Dynamic tool usage (search, DB, calculator, code, APIs, etc.)
- Maintains memory of conversation and intermediate results