Learn RAG - 25 RAG Variants Explained & Implemented

A deep dive into 25 Retrieval-Augmented Generation (RAG) variants: from foundational architectures to advanced frameworks, each with detailed tutorials and runnable Python implementations.

---

Overview

This repository is a hands-on learning resource for developers and researchers exploring RAG techniques. It covers 25 RAG variants ranging from the classic Standard RAG to advanced frameworks like Graph RAG. Each variant includes:

• In-depth tutorials (bilingual: English & Chinese): architecture analysis, working principles, use cases, and trade-off analysis
• Runnable code: complete Python implementations using sentence-transformers + FAISS + OpenAI API

---

Five Evolutionary Directions

Direction	Use Case	Representative Variants
Cost-Control	Resource-constrained environments	CORAG, Rule-based RAG
Real-time Interactive	Finance / News applications	Speculative RAG, Adaptive RAG
Domain Expert	Specialized domains	Agentic RAG, REALM
Cognitive Enhancement	Complex reasoning tasks	Self-RAG, ReAct, Graph RAG
Security & Compliance	Sensitive data handling	CRAG, EACO-RAG

---

Index of 25 RAG Variants

Foundational Architectures

#	Variant	Directory	Core Idea
01	Standard RAG	01_standard_rag	The classic retrieve-then-generate paradigm
02	CRAG (Corrective RAG)	02_crag	Iterative feedback loops to correct low-confidence outputs
03	Speculative RAG	03_speculative_rag	Dual-model strategy: specialist drafts, generalist verifies
04	Fusion RAG	04_fusion_rag	Hybrid retrieval: semantic + keyword multi-path recall
05	Agentic RAG	05_agentic_rag	Agent-driven autonomous retrieval and generation

Adaptive & Feedback

#	Variant	Directory	Core Idea
06	Self-RAG	06_self_rag	LLM self-evaluates and iteratively improves output
07	Adaptive RAG	07_adaptive_rag	Gating mechanism dynamically decides whether retrieval is needed
08	REFEED	08_refeed	Feedback-based secondary retrieval from initial output
09	REALM	09_realm	Retrieval integrated into pre-trained language models
10	RAPTOR	10_raptor	Recursive abstraction for hierarchical tree-structured retrieval

Model Integration & Enhancement

#	Variant	Directory	Core Idea
11	REVEAL	11_reveal	Multimodal visual-language retrieval augmentation
12	ReAct	12_react	Interleaved reasoning and action chain-of-thought
13	REPLUG	13_replug	Black-box retrieval plugin without model modification
14	Memo RAG	14_memo_rag	Memory-compressed retrieval clue generation
15	ATLAS	15_atlas	Retrieval fused into transformer attention layers

Architectural Innovation

#	Variant	Directory	Core Idea
16	RETRO	16_retro	Chunk-level cross-attention retrieval enhancement
17	Auto RAG	17_auto_rag	Automated iterative optimization until confidence threshold
18	CORAG	18_corag	Monte Carlo Tree Search for cost-constrained optimization
19	EACO-RAG	19_eaco_rag	Edge computing distributed retrieval
20	Rule-based RAG	20_rule_rag	Explicit rule engine for filtering and guided retrieval

Interaction & Structured Retrieval

#	Variant	Directory	Core Idea
21	Conversational RAG	21_conversational_rag	Multi-turn dialogue with persistent context retrieval
22	Iterative RAG	22_iterative_rag	Iterative loops for progressive answer refinement
23	Tree Retrieval	23_tree_retrieval	Complex query decomposition into sub-query trees
24	CRAT	24_crat	Causal reasoning-augmented translation retrieval
25	Graph RAG	25_graph_rag	Knowledge graph-driven structured retrieval

---

Quick Start

Prerequisites

• Python >= 3.12
• uv package manager

Installation

# Clone the repository
git clone https://github.com/your-username/learn-rag.git
cd learn-rag

# Install dependencies
uv sync

# Configure API Key (required for LLM generation)
cp .env.example .env
# Edit .env and add your OPENAI_API_KEY

Run an Example

# Run any variant
uv run python variants/01_standard_rag/main.py

# If you need to load environment variables from .env
source .env  # or: export OPENAI_API_KEY=sk-your-key
uv run python variants/01_standard_rag/main.py

---

Project Structure

learn-rag/
├── README.md                          # English documentation (this file)
├── README_CN.md                       # Chinese documentation
├── pyproject.toml                     # uv project configuration
├── .python-version                    # Python version
├── .env.example                       # Environment variable template
├── images/                            # Architecture diagrams
│   ├── rag_01.png
│   └── ...
└── variants/                          # 25 RAG variants
    ├── 01_standard_rag/
    │   ├── README.md                  # English tutorial
    │   ├── README_CN.md               # Chinese tutorial
    │   └── main.py                    # Runnable script
    ├── 02_crag/
    │   └── ...
    └── 25_graph_rag/
        └── ...

---

Tech Stack

• Embeddings: sentence-transformers (all-MiniLM-L6-v2)
• Vector Index: FAISS
• LLM: OpenAI API
• NLP: NLTK
• Knowledge Graph: NetworkX + spaCy

---

Choosing the Right Variant

• Getting started: Begin with Standard RAG to understand the fundamentals
• Production deployment: Consider Fusion RAG for multi-method retrieval with high recall
• Complex reasoning: Try Self-RAG or ReAct for autonomous reasoning
• Resource-constrained: Use CORAG or Rule-based RAG to control costs
• Multimodal needs: Explore REVEAL for text and visual data

---

References

• Original article: Zhihu - Baiqi
• Repost: SegmentFault - AIGC Deep Dive

---

License

MIT