Agentic Design

Patterns
🤖

Agentic RAG(AgRAG)

Autonomous retrieval-augmented generation systems with self-directed planning, retrieval, and reasoning capabilities

Complexity: highKnowledge Retrieval (RAG)

🎯 30-Second Overview

Pattern: Autonomous agent systems that dynamically plan, execute, and adapt multi-step retrieval strategies using tools and reasoning

Why: Enables complex information gathering workflows that require strategic thinking, adaptation, and multi-source synthesis

Key Insight: ReAct-style planning with tool orchestration allows agents to reason about retrieval strategies and adapt based on results

⚡ Quick Implementation

1Agent Planning:Decompose complex queries into multi-step retrieval plans with reasoning
2Tool Orchestration:Execute retrieval actions using specialized tools (search, DB, APIs)
3Dynamic Adaptation:Adapt strategy based on intermediate results and feedback
4Evidence Synthesis:Integrate multi-source evidence with reasoning chains
5Response Generation:Generate final answer with full citation and reasoning trace
Example: complex_query → plan_decomposition → [retrieve_1, retrieve_2, ...] → adapt_strategy → synthesize → response

📋 Do's & Don'ts

✅Implement explicit planning with ReAct or function calling for retrieval strategies
✅Use tool-augmented retrieval with specialized APIs, search engines, and databases
✅Enable adaptive query refinement based on intermediate retrieval results
✅Maintain detailed reasoning traces and citation provenance throughout execution
✅Implement error recovery and alternative retrieval strategies for failed attempts
❌Allow uncontrolled agent loops without termination conditions or cost limits
❌Skip validation of retrieved information quality and source credibility
❌Ignore user intent drift during multi-step retrieval processes
❌Rely solely on single retrieval modality without exploring alternatives
❌Neglect privacy and security considerations for sensitive information retrieval

🚦 When to Use

Use When

  • • Complex multi-step research requiring strategic information gathering
  • • Cross-domain queries needing diverse knowledge sources and reasoning
  • • Dynamic knowledge environments where retrieval strategies must adapt
  • • High-stakes applications requiring explainable reasoning and provenance
  • • Research and analysis tasks benefiting from human-like information seeking

Avoid When

  • • Simple factual queries adequately served by standard RAG approaches
  • • Real-time applications with strict latency and cost constraints
  • • Domains with limited tool availability or API access restrictions
  • • Applications where deterministic retrieval behavior is required
  • • Resource-constrained environments unable to support complex agent reasoning

📊 Key Metrics

Plan Execution Success
Percentage of agent plans successfully completed without errors
Retrieval Strategy Effectiveness
Quality improvement from adaptive vs fixed retrieval strategies
Evidence Synthesis Quality
Coherence and completeness of multi-source information integration
Reasoning Trace Fidelity
Accuracy and explainability of step-by-step reasoning chains
Tool Utilization Efficiency
Optimal use of available retrieval tools and APIs
Cost-Quality Tradeoff
Answer quality per unit computational cost and API calls

💡 Top Use Cases

Research Intelligence: Scientific literature review with adaptive search strategies and cross-reference validation
Business Analysis: Market research combining financial data, news, and industry reports with strategic reasoning
Legal Discovery: Case law research with multi-jurisdictional search and precedent analysis workflows
Technical Investigation: Software debugging and troubleshooting with dynamic tool selection and reasoning
Medical Diagnosis Support: Clinical decision support with evidence gathering from multiple medical databases

References & Further Reading

Deepen your understanding with these curated resources

Foundational Papers & Agentic RAG Research

ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2022)

Toolformer: Language Models Can Teach Themselves to Use Tools (Schick et al., 2023)

HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face (Shen et al., 2023)

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., 2020)

Agent Frameworks & Planning Methods

AutoGPT: An Autonomous GPT-4 Experiment

LangGraph: Building Stateful, Multi-Actor Applications with LLMs

CrewAI: Framework for Orchestrating Role-Playing, Autonomous AI Agents

Microsoft AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

Tool Integration & Function Calling

OpenAI Function Calling: Connecting GPTs to External Tools

Anthropic Tool Use and Function Calling Guide

Google Gemini Function Calling and Tool Integration

LangChain Tools and Agents: Comprehensive Implementation Guide

Multi-Step Reasoning & Planning

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022)

Tree of Thoughts: Deliberate Problem Solving with Large Language Models (Yao et al., 2023)

Self-Consistency Improves Chain of Thought Reasoning (Wang et al., 2022)

Program-aided Language Models (Gao et al., 2023)

Implementation Frameworks & Production Systems

LlamaIndex Agent Framework: Tool-Augmented RAG Systems

Haystack Agent Pipeline: Multi-Step Information Retrieval

Semantic Kernel: AI Orchestration Framework by Microsoft

LangSmith: Production Monitoring for Agent-Based RAG Systems

Evaluation & Quality Assessment

AgentBench: Evaluating LLMs as Agents (Liu et al., 2023)

WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents (Yao et al., 2022)

ToolQA: A Dataset for LLM Question Answering with External Tools (Zhuang et al., 2023)

RAGAS: Automated Evaluation Framework for RAG Applications

Contribute to this collection

Know a great resource? Submit a pull request to add it.

Contribute

Patterns

closed

Loading...